P. Mohajerin Esfahani
info
Please Note
<p>This page displays the records of the person named above and is not linked to a unique person identifier. This record may need to be merged to a profile.</p>
53 records found
1
Constraint-Based Structural Attention in Graph Transformers
An Inverse-Optimisation-Based Approach
Many problems in science and engineering involve data that is naturally graph-structured: molecules, proteins, knowledge bases, and road networks. Graph Transformers have become one of the leading model families for this kind of data. Because plain self-attention treats its input as an unordered set, graph topology (edges, distances, and connectivity patterns) must be supplied to the model explicitly, either through structural encodings added to the node features or, more directly, through a learned bias added to the attention logits. In the latter case, a structural term is added to the content score, where it competes with that score on equal footing and applies the same correction at every layer, regardless of whether content attention already respects the structural prior.
We take a different starting point. Softmax attention can itself be written as the solution to an entropy-regularized optimization problem, which suggests treating structure not as another term in the objective but as a constraint on the attention itself. This thesis develops that idea into a Graph Transformer in which structural priors are imposed as per-pair inequality constraints on the attention distribution. Each correction is governed by a Lagrangian dual variable that the model carries across transformer layers, growing for pairs where content attention falls short of the structural prior and relaxing where the prior is already met.
The result is a Graph Transformer that learns structural constraints within its attention mechanism rather than treating structure as a fixed additive term. Because the dual variable reacts to how far the realized attention falls short of the structural target, the correction is active only where structure is genuinely underserved and silent elsewhere. The construction is agnostic to which pairwise graph signal serves as the constraint, and multiple signals can be combined without architectural changes. On the ZINC and LRGB Peptides molecular benchmarks, the adaptive constraint mechanism consistently improves over the matched baseline, in which the dual is held fixed at unit scale. ...
We take a different starting point. Softmax attention can itself be written as the solution to an entropy-regularized optimization problem, which suggests treating structure not as another term in the objective but as a constraint on the attention itself. This thesis develops that idea into a Graph Transformer in which structural priors are imposed as per-pair inequality constraints on the attention distribution. Each correction is governed by a Lagrangian dual variable that the model carries across transformer layers, growing for pairs where content attention falls short of the structural prior and relaxing where the prior is already met.
The result is a Graph Transformer that learns structural constraints within its attention mechanism rather than treating structure as a fixed additive term. Because the dual variable reacts to how far the realized attention falls short of the structural target, the correction is active only where structure is genuinely underserved and silent elsewhere. The construction is agnostic to which pairwise graph signal serves as the constraint, and multiple signals can be combined without architectural changes. On the ZINC and LRGB Peptides molecular benchmarks, the adaptive constraint mechanism consistently improves over the matched baseline, in which the dual is held fixed at unit scale. ...
Many problems in science and engineering involve data that is naturally graph-structured: molecules, proteins, knowledge bases, and road networks. Graph Transformers have become one of the leading model families for this kind of data. Because plain self-attention treats its input as an unordered set, graph topology (edges, distances, and connectivity patterns) must be supplied to the model explicitly, either through structural encodings added to the node features or, more directly, through a learned bias added to the attention logits. In the latter case, a structural term is added to the content score, where it competes with that score on equal footing and applies the same correction at every layer, regardless of whether content attention already respects the structural prior.
We take a different starting point. Softmax attention can itself be written as the solution to an entropy-regularized optimization problem, which suggests treating structure not as another term in the objective but as a constraint on the attention itself. This thesis develops that idea into a Graph Transformer in which structural priors are imposed as per-pair inequality constraints on the attention distribution. Each correction is governed by a Lagrangian dual variable that the model carries across transformer layers, growing for pairs where content attention falls short of the structural prior and relaxing where the prior is already met.
The result is a Graph Transformer that learns structural constraints within its attention mechanism rather than treating structure as a fixed additive term. Because the dual variable reacts to how far the realized attention falls short of the structural target, the correction is active only where structure is genuinely underserved and silent elsewhere. The construction is agnostic to which pairwise graph signal serves as the constraint, and multiple signals can be combined without architectural changes. On the ZINC and LRGB Peptides molecular benchmarks, the adaptive constraint mechanism consistently improves over the matched baseline, in which the dual is held fixed at unit scale.
We take a different starting point. Softmax attention can itself be written as the solution to an entropy-regularized optimization problem, which suggests treating structure not as another term in the objective but as a constraint on the attention itself. This thesis develops that idea into a Graph Transformer in which structural priors are imposed as per-pair inequality constraints on the attention distribution. Each correction is governed by a Lagrangian dual variable that the model carries across transformer layers, growing for pairs where content attention falls short of the structural prior and relaxing where the prior is already met.
The result is a Graph Transformer that learns structural constraints within its attention mechanism rather than treating structure as a fixed additive term. Because the dual variable reacts to how far the realized attention falls short of the structural target, the correction is active only where structure is genuinely underserved and silent elsewhere. The construction is agnostic to which pairwise graph signal serves as the constraint, and multiple signals can be combined without architectural changes. On the ZINC and LRGB Peptides molecular benchmarks, the adaptive constraint mechanism consistently improves over the matched baseline, in which the dual is held fixed at unit scale.
This thesis addresses the challenge of autonomous bathymetric mapping, the process of creat-ing topographical maps of the ocean floor, using an Autonomous Underwater Vehicle (AUV).Conventional surveys rely on expensive crewed vessels and often fail to exploit the underlyingstatistical patterns of the collected bathymetric data. This work proposes a model-drivenestimation framework that treats bathymetric mapping as a nonlinear system identification problem.
Using properties of the underlying state-space model and by representing the seabed through alinear combination of known basis functions, a novel estimation algorithm is derived. The pro-posed method formulates a dual Bayesian state-parameter estimator in which affine MinimumMean Square Error (MMSE) estimators are constructed for both the unknown model param-eters and the latent system states. By alternating between these closed-form estimators ina fixed-point iteration, the algorithm progressively reduces state-induced uncertainty andimproves parameter estimation accuracy.
Finally, a filtering extension of the dual state-parameter estimator is introduced. This exten-sion enables scalable processing of large data sets by operating online, while incurring a limitedloss in accuracy compared to the batch formulation. The proposed estimators are evaluatedin Monte Carlo experiments and benchmarked against a state-of-the-art methods. The resultsshow that the proposed methods consistently outperform existing estimation algorithms interms of accuracy and computational complexity. ...
Using properties of the underlying state-space model and by representing the seabed through alinear combination of known basis functions, a novel estimation algorithm is derived. The pro-posed method formulates a dual Bayesian state-parameter estimator in which affine MinimumMean Square Error (MMSE) estimators are constructed for both the unknown model param-eters and the latent system states. By alternating between these closed-form estimators ina fixed-point iteration, the algorithm progressively reduces state-induced uncertainty andimproves parameter estimation accuracy.
Finally, a filtering extension of the dual state-parameter estimator is introduced. This exten-sion enables scalable processing of large data sets by operating online, while incurring a limitedloss in accuracy compared to the batch formulation. The proposed estimators are evaluatedin Monte Carlo experiments and benchmarked against a state-of-the-art methods. The resultsshow that the proposed methods consistently outperform existing estimation algorithms interms of accuracy and computational complexity. ...
This thesis addresses the challenge of autonomous bathymetric mapping, the process of creat-ing topographical maps of the ocean floor, using an Autonomous Underwater Vehicle (AUV).Conventional surveys rely on expensive crewed vessels and often fail to exploit the underlyingstatistical patterns of the collected bathymetric data. This work proposes a model-drivenestimation framework that treats bathymetric mapping as a nonlinear system identification problem.
Using properties of the underlying state-space model and by representing the seabed through alinear combination of known basis functions, a novel estimation algorithm is derived. The pro-posed method formulates a dual Bayesian state-parameter estimator in which affine MinimumMean Square Error (MMSE) estimators are constructed for both the unknown model param-eters and the latent system states. By alternating between these closed-form estimators ina fixed-point iteration, the algorithm progressively reduces state-induced uncertainty andimproves parameter estimation accuracy.
Finally, a filtering extension of the dual state-parameter estimator is introduced. This exten-sion enables scalable processing of large data sets by operating online, while incurring a limitedloss in accuracy compared to the batch formulation. The proposed estimators are evaluatedin Monte Carlo experiments and benchmarked against a state-of-the-art methods. The resultsshow that the proposed methods consistently outperform existing estimation algorithms interms of accuracy and computational complexity.
Using properties of the underlying state-space model and by representing the seabed through alinear combination of known basis functions, a novel estimation algorithm is derived. The pro-posed method formulates a dual Bayesian state-parameter estimator in which affine MinimumMean Square Error (MMSE) estimators are constructed for both the unknown model param-eters and the latent system states. By alternating between these closed-form estimators ina fixed-point iteration, the algorithm progressively reduces state-induced uncertainty andimproves parameter estimation accuracy.
Finally, a filtering extension of the dual state-parameter estimator is introduced. This exten-sion enables scalable processing of large data sets by operating online, while incurring a limitedloss in accuracy compared to the batch formulation. The proposed estimators are evaluatedin Monte Carlo experiments and benchmarked against a state-of-the-art methods. The resultsshow that the proposed methods consistently outperform existing estimation algorithms interms of accuracy and computational complexity.
This thesis develops high-performance numerical methods for convex optimization, variational inequalities, and game theory, targeting computational bottlenecks in modern large-scale systems. By leveraging the underlying mathematical structure of these problems, this work bridges the gap between abstract operator theory and real-time control and strategic decision-making applications.
The first core contribution focuses on accelerating first-order methods for smooth and nonsmooth convex optimization. We introduce adaptive step-size rules and coupled smoothing–momentum techniques that achieve optimal convergence rates. These methods are designed to exploit problem structure, ensuring computational efficiency and enabling fast convergence without requiring prior knowledge of global problem parameters.
Extending beyond single-agent optimization, the research adopts the framework of variational inequalities to address complex equilibrium problems. We propose projection-free algorithms and specialized splitting methods for settings in which traditional projection operators are computationally expensive. This unified approach enables efficient computation of equilibria in dynamic games and distributionally robust models, where decision-makers must account for both strategic interactions and data uncertainty.
The practical relevance of these developments is demonstrated through real-world applications and the introduction of an open-source computational toolkit. Collectively, these contributions provide a scalable and robust framework for fast, structure-aware decision-making in complex multi-agent systems. ...
The first core contribution focuses on accelerating first-order methods for smooth and nonsmooth convex optimization. We introduce adaptive step-size rules and coupled smoothing–momentum techniques that achieve optimal convergence rates. These methods are designed to exploit problem structure, ensuring computational efficiency and enabling fast convergence without requiring prior knowledge of global problem parameters.
Extending beyond single-agent optimization, the research adopts the framework of variational inequalities to address complex equilibrium problems. We propose projection-free algorithms and specialized splitting methods for settings in which traditional projection operators are computationally expensive. This unified approach enables efficient computation of equilibria in dynamic games and distributionally robust models, where decision-makers must account for both strategic interactions and data uncertainty.
The practical relevance of these developments is demonstrated through real-world applications and the introduction of an open-source computational toolkit. Collectively, these contributions provide a scalable and robust framework for fast, structure-aware decision-making in complex multi-agent systems. ...
This thesis develops high-performance numerical methods for convex optimization, variational inequalities, and game theory, targeting computational bottlenecks in modern large-scale systems. By leveraging the underlying mathematical structure of these problems, this work bridges the gap between abstract operator theory and real-time control and strategic decision-making applications.
The first core contribution focuses on accelerating first-order methods for smooth and nonsmooth convex optimization. We introduce adaptive step-size rules and coupled smoothing–momentum techniques that achieve optimal convergence rates. These methods are designed to exploit problem structure, ensuring computational efficiency and enabling fast convergence without requiring prior knowledge of global problem parameters.
Extending beyond single-agent optimization, the research adopts the framework of variational inequalities to address complex equilibrium problems. We propose projection-free algorithms and specialized splitting methods for settings in which traditional projection operators are computationally expensive. This unified approach enables efficient computation of equilibria in dynamic games and distributionally robust models, where decision-makers must account for both strategic interactions and data uncertainty.
The practical relevance of these developments is demonstrated through real-world applications and the introduction of an open-source computational toolkit. Collectively, these contributions provide a scalable and robust framework for fast, structure-aware decision-making in complex multi-agent systems.
The first core contribution focuses on accelerating first-order methods for smooth and nonsmooth convex optimization. We introduce adaptive step-size rules and coupled smoothing–momentum techniques that achieve optimal convergence rates. These methods are designed to exploit problem structure, ensuring computational efficiency and enabling fast convergence without requiring prior knowledge of global problem parameters.
Extending beyond single-agent optimization, the research adopts the framework of variational inequalities to address complex equilibrium problems. We propose projection-free algorithms and specialized splitting methods for settings in which traditional projection operators are computationally expensive. This unified approach enables efficient computation of equilibria in dynamic games and distributionally robust models, where decision-makers must account for both strategic interactions and data uncertainty.
The practical relevance of these developments is demonstrated through real-world applications and the introduction of an open-source computational toolkit. Collectively, these contributions provide a scalable and robust framework for fast, structure-aware decision-making in complex multi-agent systems.
Towards Dynamic Inverse Optimization
A Data Aggregation Approach
Data-driven Inverse Optimization (IO) is a form of Supervised Learning where it is assumed that the output data is found by means of an optimization problem that depends on the input data. IO uses this data to approximate the optimization problem as best as possible. In the case where one wants to emulate an expert operating in a dynamic environment, the dataset obtained by measuring the expert is often contained in a small, optimal part of the total state space of the environment. When a model trained on this data finds itself in a different part of the state space, it can behave erratically.
In this thesis, we will combine Inverse Optimization with the active learning method of Dataset Aggregation (DAgger) to test if this improves model performance in dynamic settings. DAgger is an iterative process where the system is steered by the learner, creating new input data for the expert to find the best actions. This new data is then used to train a new model.
Furthermore, we propose a new algorithm, fast-DAgger, that should converge faster than the DAgger algorithm, at the possible cost of performance in the final model.
IO models trained with the DAgger and fast-DAgger algorithms are tested and compared to IO models trained on static datasets. This is done for two case studies: the Dynamic Vehicle Routing Problem as proposed by the EURO meets Neurips 2022 Vehicle Routing Competition, and the game of Tetris.
Results show the potential of combining IO with DAgger. However, DAgger is not always better than training with a static dataset. DAgger can only be helpful when the static training data is limited to a part of the total state space and when this data does not generalize well to the total state space. The fast-DAgger algorithm did not show a significant speed-up compared to the normal DAgger algorithm in the case studies. However, this is very dependent on the specifics of the model and the hyperparameters of the DAgger algorithm.
...
In this thesis, we will combine Inverse Optimization with the active learning method of Dataset Aggregation (DAgger) to test if this improves model performance in dynamic settings. DAgger is an iterative process where the system is steered by the learner, creating new input data for the expert to find the best actions. This new data is then used to train a new model.
Furthermore, we propose a new algorithm, fast-DAgger, that should converge faster than the DAgger algorithm, at the possible cost of performance in the final model.
IO models trained with the DAgger and fast-DAgger algorithms are tested and compared to IO models trained on static datasets. This is done for two case studies: the Dynamic Vehicle Routing Problem as proposed by the EURO meets Neurips 2022 Vehicle Routing Competition, and the game of Tetris.
Results show the potential of combining IO with DAgger. However, DAgger is not always better than training with a static dataset. DAgger can only be helpful when the static training data is limited to a part of the total state space and when this data does not generalize well to the total state space. The fast-DAgger algorithm did not show a significant speed-up compared to the normal DAgger algorithm in the case studies. However, this is very dependent on the specifics of the model and the hyperparameters of the DAgger algorithm.
...
Data-driven Inverse Optimization (IO) is a form of Supervised Learning where it is assumed that the output data is found by means of an optimization problem that depends on the input data. IO uses this data to approximate the optimization problem as best as possible. In the case where one wants to emulate an expert operating in a dynamic environment, the dataset obtained by measuring the expert is often contained in a small, optimal part of the total state space of the environment. When a model trained on this data finds itself in a different part of the state space, it can behave erratically.
In this thesis, we will combine Inverse Optimization with the active learning method of Dataset Aggregation (DAgger) to test if this improves model performance in dynamic settings. DAgger is an iterative process where the system is steered by the learner, creating new input data for the expert to find the best actions. This new data is then used to train a new model.
Furthermore, we propose a new algorithm, fast-DAgger, that should converge faster than the DAgger algorithm, at the possible cost of performance in the final model.
IO models trained with the DAgger and fast-DAgger algorithms are tested and compared to IO models trained on static datasets. This is done for two case studies: the Dynamic Vehicle Routing Problem as proposed by the EURO meets Neurips 2022 Vehicle Routing Competition, and the game of Tetris.
Results show the potential of combining IO with DAgger. However, DAgger is not always better than training with a static dataset. DAgger can only be helpful when the static training data is limited to a part of the total state space and when this data does not generalize well to the total state space. The fast-DAgger algorithm did not show a significant speed-up compared to the normal DAgger algorithm in the case studies. However, this is very dependent on the specifics of the model and the hyperparameters of the DAgger algorithm.
In this thesis, we will combine Inverse Optimization with the active learning method of Dataset Aggregation (DAgger) to test if this improves model performance in dynamic settings. DAgger is an iterative process where the system is steered by the learner, creating new input data for the expert to find the best actions. This new data is then used to train a new model.
Furthermore, we propose a new algorithm, fast-DAgger, that should converge faster than the DAgger algorithm, at the possible cost of performance in the final model.
IO models trained with the DAgger and fast-DAgger algorithms are tested and compared to IO models trained on static datasets. This is done for two case studies: the Dynamic Vehicle Routing Problem as proposed by the EURO meets Neurips 2022 Vehicle Routing Competition, and the game of Tetris.
Results show the potential of combining IO with DAgger. However, DAgger is not always better than training with a static dataset. DAgger can only be helpful when the static training data is limited to a part of the total state space and when this data does not generalize well to the total state space. The fast-DAgger algorithm did not show a significant speed-up compared to the normal DAgger algorithm in the case studies. However, this is very dependent on the specifics of the model and the hyperparameters of the DAgger algorithm.
Fault Diagnosis and Estimation in Inkjet Printers Using Self-Sensing Piezo Actuators
1- Introduction 2- Problem Statement 3- PART I – Fault Detection and Isolation (FDI) 4- PART II – Fault Estimation (FE) 5- Summary of Key Contributions 6- Conclusions and Future Work
Diagnosing nozzle faults in high-end industrial printers, such as the ones developed by Canon Production Printing (CPP), remains challenging due to the interplay of fluid dynamics and mechanical actuation. These systems rely on self-sensing signals that are often subtle and nonlinear, complicating both detection and interpretation. However, accurate and timely diagnosis is essential to maintain print quality, minimize waste, and reduce maintenance effort. This thesis investigates hybrid fault diagnosis methods that integrate model-based and data-driven techniques to improve detection reliability and generalization, particularly for piezoelectric inkjet systems. Traditional fault detection approaches in this context often rely on rule-based thresholds applied to features extracted from self-sensing signals. Although these methods can be effective, they are typically sensitive to variations in operating conditions. In contrast, model-based techniques use simplified system dynamics to generate residual signals that reflect deviations from expected behavior. In this thesis, we propose a hybrid framework that addresses the Fault Detection and Isolation (FDI) problem from a frequency domain perspective. By learning from signal characteristics, the method avoids the need for manually defined thresholds and predefined reference signatures. Instead, it uses classifiers trained to distinguish between different fault types and improve the adaptability to unseen cases. Building on this framework, the second part of the thesis addresses Fault Estimation (FE), aiming to reconstruct how faults evolve over time. A linear model-based estimation scheme is developed in both discrete-time and continuous-time forms. Even though this approach simplifies certain nonlinear dynamics, it provides useful fault tracking results, particularly for moderate fault levels. The evaluation on synthetic datasets shows that the proposed FDI and FE methods offer interpretable and reasonably accurate results. However, challenges remain when applied to physics-based data, particularly due to nonlinear effects, variable initial conditions, and numerical sensitivity.
...
Diagnosing nozzle faults in high-end industrial printers, such as the ones developed by Canon Production Printing (CPP), remains challenging due to the interplay of fluid dynamics and mechanical actuation. These systems rely on self-sensing signals that are often subtle and nonlinear, complicating both detection and interpretation. However, accurate and timely diagnosis is essential to maintain print quality, minimize waste, and reduce maintenance effort. This thesis investigates hybrid fault diagnosis methods that integrate model-based and data-driven techniques to improve detection reliability and generalization, particularly for piezoelectric inkjet systems. Traditional fault detection approaches in this context often rely on rule-based thresholds applied to features extracted from self-sensing signals. Although these methods can be effective, they are typically sensitive to variations in operating conditions. In contrast, model-based techniques use simplified system dynamics to generate residual signals that reflect deviations from expected behavior. In this thesis, we propose a hybrid framework that addresses the Fault Detection and Isolation (FDI) problem from a frequency domain perspective. By learning from signal characteristics, the method avoids the need for manually defined thresholds and predefined reference signatures. Instead, it uses classifiers trained to distinguish between different fault types and improve the adaptability to unseen cases. Building on this framework, the second part of the thesis addresses Fault Estimation (FE), aiming to reconstruct how faults evolve over time. A linear model-based estimation scheme is developed in both discrete-time and continuous-time forms. Even though this approach simplifies certain nonlinear dynamics, it provides useful fault tracking results, particularly for moderate fault levels. The evaluation on synthetic datasets shows that the proposed FDI and FE methods offer interpretable and reasonably accurate results. However, challenges remain when applied to physics-based data, particularly due to nonlinear effects, variable initial conditions, and numerical sensitivity.
Master thesis
(2024)
-
M.R. Delgado Gosalvez, Peyman Mohajerin Esfahani, Steven Wilkins, Mohammad Boveiri, Shengling Shi
The European Union (EU) has set the goal to have a carbon-neutral economy by 2050. To achieve this, a key sector to focus on is the transportation sector. It will be especially challenging though to decarbonise the larger vehicles from the transportation sector, the trucks, ships and airplanes.
The typical way to respond to this or any other design challenge is by making use of the top-down/sequential design process, where first the larger parts of the design are established, and later the finer details are developed. However such an approach cannot guarantee that the design at the end is optimal or anywhere close to it. This makes the sequential design approach ill-suited to tackle the challenge of decarbonising the EU transportation sector. Co-Design on the other hand, a simultaneous design approach, can guarantee optimality, and would be a much more appropriate approach if it wasn’t for it being limited to only consider around 10-20 components.
The work in this report first investigated where this limitation of 10-20 components comes from. It is described how the source of the problem does not originate from topology optimisation, where most research has gone into, but rather it originates from the vast number of isomorphic topologies that are created during the topology generation process. Adding more data to describe the components only increases the number of isomorphic topologies even further. Since the results of the topology generation process are fed to the topology optimisation process, the entire Co-Design process is impacted by these isomorphic topologies.
The work that follows therefore focused on topology generation for Co-Design by testing a candidate method that allows topologies to be generated without any of the isomorphic topologies while describing components in more detail than what has been done in literature. The approach taken here relied, unlike what has been done so far, on the use of adjacency matrices and the interconnected system model. With this, new insight into topology generation for Co-Design was developed. The use of adjacency matrices were also key to automate the formation of constraints, which allowed to generate topologies for any set of components, also a first in the literature. The program TopoGen was developed to make this possible. The results that were obtained showed how using requirement chains and ordering instances, at least in simple cases, the 10-20 components limitation is fully overcome. This confirmed that the presence of isomorphic topologies the cause was for 10-20 components limitation in Co-Design. Still, for more complex cases, mistakes were present. However it was also shown how these too may be prevented in the future. This means that with further research and development, the candidate method presented here can be adjusted so that Co-Design can be applied even to complex cases, and therefore be used to develop solutions needed to decarbonise the EU transportation sector.
...
The typical way to respond to this or any other design challenge is by making use of the top-down/sequential design process, where first the larger parts of the design are established, and later the finer details are developed. However such an approach cannot guarantee that the design at the end is optimal or anywhere close to it. This makes the sequential design approach ill-suited to tackle the challenge of decarbonising the EU transportation sector. Co-Design on the other hand, a simultaneous design approach, can guarantee optimality, and would be a much more appropriate approach if it wasn’t for it being limited to only consider around 10-20 components.
The work in this report first investigated where this limitation of 10-20 components comes from. It is described how the source of the problem does not originate from topology optimisation, where most research has gone into, but rather it originates from the vast number of isomorphic topologies that are created during the topology generation process. Adding more data to describe the components only increases the number of isomorphic topologies even further. Since the results of the topology generation process are fed to the topology optimisation process, the entire Co-Design process is impacted by these isomorphic topologies.
The work that follows therefore focused on topology generation for Co-Design by testing a candidate method that allows topologies to be generated without any of the isomorphic topologies while describing components in more detail than what has been done in literature. The approach taken here relied, unlike what has been done so far, on the use of adjacency matrices and the interconnected system model. With this, new insight into topology generation for Co-Design was developed. The use of adjacency matrices were also key to automate the formation of constraints, which allowed to generate topologies for any set of components, also a first in the literature. The program TopoGen was developed to make this possible. The results that were obtained showed how using requirement chains and ordering instances, at least in simple cases, the 10-20 components limitation is fully overcome. This confirmed that the presence of isomorphic topologies the cause was for 10-20 components limitation in Co-Design. Still, for more complex cases, mistakes were present. However it was also shown how these too may be prevented in the future. This means that with further research and development, the candidate method presented here can be adjusted so that Co-Design can be applied even to complex cases, and therefore be used to develop solutions needed to decarbonise the EU transportation sector.
...
The European Union (EU) has set the goal to have a carbon-neutral economy by 2050. To achieve this, a key sector to focus on is the transportation sector. It will be especially challenging though to decarbonise the larger vehicles from the transportation sector, the trucks, ships and airplanes.
The typical way to respond to this or any other design challenge is by making use of the top-down/sequential design process, where first the larger parts of the design are established, and later the finer details are developed. However such an approach cannot guarantee that the design at the end is optimal or anywhere close to it. This makes the sequential design approach ill-suited to tackle the challenge of decarbonising the EU transportation sector. Co-Design on the other hand, a simultaneous design approach, can guarantee optimality, and would be a much more appropriate approach if it wasn’t for it being limited to only consider around 10-20 components.
The work in this report first investigated where this limitation of 10-20 components comes from. It is described how the source of the problem does not originate from topology optimisation, where most research has gone into, but rather it originates from the vast number of isomorphic topologies that are created during the topology generation process. Adding more data to describe the components only increases the number of isomorphic topologies even further. Since the results of the topology generation process are fed to the topology optimisation process, the entire Co-Design process is impacted by these isomorphic topologies.
The work that follows therefore focused on topology generation for Co-Design by testing a candidate method that allows topologies to be generated without any of the isomorphic topologies while describing components in more detail than what has been done in literature. The approach taken here relied, unlike what has been done so far, on the use of adjacency matrices and the interconnected system model. With this, new insight into topology generation for Co-Design was developed. The use of adjacency matrices were also key to automate the formation of constraints, which allowed to generate topologies for any set of components, also a first in the literature. The program TopoGen was developed to make this possible. The results that were obtained showed how using requirement chains and ordering instances, at least in simple cases, the 10-20 components limitation is fully overcome. This confirmed that the presence of isomorphic topologies the cause was for 10-20 components limitation in Co-Design. Still, for more complex cases, mistakes were present. However it was also shown how these too may be prevented in the future. This means that with further research and development, the candidate method presented here can be adjusted so that Co-Design can be applied even to complex cases, and therefore be used to develop solutions needed to decarbonise the EU transportation sector.
The typical way to respond to this or any other design challenge is by making use of the top-down/sequential design process, where first the larger parts of the design are established, and later the finer details are developed. However such an approach cannot guarantee that the design at the end is optimal or anywhere close to it. This makes the sequential design approach ill-suited to tackle the challenge of decarbonising the EU transportation sector. Co-Design on the other hand, a simultaneous design approach, can guarantee optimality, and would be a much more appropriate approach if it wasn’t for it being limited to only consider around 10-20 components.
The work in this report first investigated where this limitation of 10-20 components comes from. It is described how the source of the problem does not originate from topology optimisation, where most research has gone into, but rather it originates from the vast number of isomorphic topologies that are created during the topology generation process. Adding more data to describe the components only increases the number of isomorphic topologies even further. Since the results of the topology generation process are fed to the topology optimisation process, the entire Co-Design process is impacted by these isomorphic topologies.
The work that follows therefore focused on topology generation for Co-Design by testing a candidate method that allows topologies to be generated without any of the isomorphic topologies while describing components in more detail than what has been done in literature. The approach taken here relied, unlike what has been done so far, on the use of adjacency matrices and the interconnected system model. With this, new insight into topology generation for Co-Design was developed. The use of adjacency matrices were also key to automate the formation of constraints, which allowed to generate topologies for any set of components, also a first in the literature. The program TopoGen was developed to make this possible. The results that were obtained showed how using requirement chains and ordering instances, at least in simple cases, the 10-20 components limitation is fully overcome. This confirmed that the presence of isomorphic topologies the cause was for 10-20 components limitation in Co-Design. Still, for more complex cases, mistakes were present. However it was also shown how these too may be prevented in the future. This means that with further research and development, the candidate method presented here can be adjusted so that Co-Design can be applied even to complex cases, and therefore be used to develop solutions needed to decarbonise the EU transportation sector.
Fault Detection and Isolation for High-End Industrial Printers
A hybrid model- and data-based approach
Master thesis
(2024)
-
C.D. van Peijpe, P. Mohajerin Esfahani, Farhad Ghanipoor Ghanipoor, Youri de Loore, Pim Hacking
This report presents a hybrid model- and data-based method for detecting and isolating faults in the ink channel of a printer using the self-sensing capability of piezo actuators. Grey-box system identification is used to identify the parameters of the model of the ink channel. The model is used to construct the fault detection (FD) filter. The FD filter uses the piezo self-sensing signal of the printer as an input and puts out a residual signal, which is approximately zero when the system is healthy. Several methods to design the FD filter denominator are proposed. If the energy of the residual exceeds a threshold, the system is detected as faulty. The fault isolation (FI) filter uses linear regression, utilizing the residuals of the FD filter, to generate a probability vector. The entries of this vector correspond to the possible faults, and the highest entry is used to isolate the fault. Both filters are tested for a simulated and experimental dataset. For both datasets, the FD filter is shown to perform appropriately, as does the FI filter for experimental data. For simulated data, the FI filter is compared to other methods. The FI filter performs best when only regarding isolations with a high certainty.
...
This report presents a hybrid model- and data-based method for detecting and isolating faults in the ink channel of a printer using the self-sensing capability of piezo actuators. Grey-box system identification is used to identify the parameters of the model of the ink channel. The model is used to construct the fault detection (FD) filter. The FD filter uses the piezo self-sensing signal of the printer as an input and puts out a residual signal, which is approximately zero when the system is healthy. Several methods to design the FD filter denominator are proposed. If the energy of the residual exceeds a threshold, the system is detected as faulty. The fault isolation (FI) filter uses linear regression, utilizing the residuals of the FD filter, to generate a probability vector. The entries of this vector correspond to the possible faults, and the highest entry is used to isolate the fault. Both filters are tested for a simulated and experimental dataset. For both datasets, the FD filter is shown to perform appropriately, as does the FI filter for experimental data. For simulated data, the FI filter is compared to other methods. The FI filter performs best when only regarding isolations with a high certainty.
Data-Driven Optimal Control
An Inverse Optimization Model and Algorithm
In Inverse Optimization (IO), it is hypothesized that experts, when making decisions, implicitly engage in solving an optimization problem. If we can reconstruct this optimization problem using the decision data of the expert, then the behavior of the expert can be emulated. In this thesis, a novel inverse optimization model, Kernel Inverse Optimization Machine (KIOM), is proposed, utilizing kernel methods. Because its parameter space can be potentially infinite-dimensional, the model exhibits strong representation and generalization capabilities. Furthermore, empirical evidence is presented demonstrating the model’s ability to learn complex MuJoCo continuous control tasks. Subsequently, an algorithm for training KIOM, Sequential Selection Optimization (SSO), is proposed to address memory issues. SSO is a coordinate descent-based algorithm, and its memory requirements are nearly equal to the memory needed for solving one of its subproblems. Experimental results demonstrate that SSO converges to the optimal solution within a small number of iterations, highlighting its
efficiency. ...
efficiency. ...
In Inverse Optimization (IO), it is hypothesized that experts, when making decisions, implicitly engage in solving an optimization problem. If we can reconstruct this optimization problem using the decision data of the expert, then the behavior of the expert can be emulated. In this thesis, a novel inverse optimization model, Kernel Inverse Optimization Machine (KIOM), is proposed, utilizing kernel methods. Because its parameter space can be potentially infinite-dimensional, the model exhibits strong representation and generalization capabilities. Furthermore, empirical evidence is presented demonstrating the model’s ability to learn complex MuJoCo continuous control tasks. Subsequently, an algorithm for training KIOM, Sequential Selection Optimization (SSO), is proposed to address memory issues. SSO is a coordinate descent-based algorithm, and its memory requirements are nearly equal to the memory needed for solving one of its subproblems. Experimental results demonstrate that SSO converges to the optimal solution within a small number of iterations, highlighting its
efficiency.
efficiency.
Slag Basicity Control under Ambiguity in HIsarna
Distributional Robust Control
Master thesis
(2024)
-
C. Wang, P. Mohajerin Esfahani, G.F. Max, E. Feenstra, G. de Albuquerque Gleizer
The steel industry is one of the largest emitters of greenhouse gases. Therefore, there is a need to develop revolutionary sustainable methods for producing iron and steel. HIsarna is one such sustainable method developed by TATA Steel Europe for a long time. Due to the flexibility offered by this iron-making process which allows using unprocessed iron ore, it is a promising technology. Currently, work is going on to stabilize and optimize this iron-making method. One of the steps of that process is maintaining the optimal level of slag in terms of its chemical composition. It is important to regulate the slag basicity to maintain the quality of iron produced. This is where the concept of distributionally robust control comes into use, as the fluctuations in the slag basicity are random that we wish to control under imperfect knowledge of the distribution of these disturbances. The aim of this project is twofold. First, the existing controller (from previous work done on HIsarna) is implemented on the real system. This controller outperformed human operators in a simulation environment which motivates this step. Second, using techniques from distributionally robust control, improves the robustness and performance of the controller. While the existing controller was trained in a simulation environment, the uncertainty in the real system may be different from that of the simulator. Using actual measurements together with more sophisticated training may lead to a controller that can handle various material properties and operating ranges appearing more common in production.
...
The steel industry is one of the largest emitters of greenhouse gases. Therefore, there is a need to develop revolutionary sustainable methods for producing iron and steel. HIsarna is one such sustainable method developed by TATA Steel Europe for a long time. Due to the flexibility offered by this iron-making process which allows using unprocessed iron ore, it is a promising technology. Currently, work is going on to stabilize and optimize this iron-making method. One of the steps of that process is maintaining the optimal level of slag in terms of its chemical composition. It is important to regulate the slag basicity to maintain the quality of iron produced. This is where the concept of distributionally robust control comes into use, as the fluctuations in the slag basicity are random that we wish to control under imperfect knowledge of the distribution of these disturbances. The aim of this project is twofold. First, the existing controller (from previous work done on HIsarna) is implemented on the real system. This controller outperformed human operators in a simulation environment which motivates this step. Second, using techniques from distributionally robust control, improves the robustness and performance of the controller. While the existing controller was trained in a simulation environment, the uncertainty in the real system may be different from that of the simulator. Using actual measurements together with more sophisticated training may lead to a controller that can handle various material properties and operating ranges appearing more common in production.
Master thesis
(2024)
-
D. Jansen, P. Mohajerin Esfahani, A. S. Vora, A. Sharifi Kolarijani, Jens van de Graaf
The rise of social media has significantly influenced digital marketing, especially within the IT Channel Industry. Channext, a key player in automating channel marketing, faces the challenge of achieving optimal through-partner social engagement, which includes maximizing shares, impressions, clicks, likes, and comments on social media posts. However, the current approach is hindered by content overload, leading to scheduling conflicts and reduced engagement.
To address this issue, this project introduces an advanced model for the vendor-partner network, with several key contributions: • Predictive Analytics: A robust predictor is developed to capture the relationship between historical impressions data and social post characteristics, aiming to forecast future impressions with improved accuracy. • Optimized Scheduling: Predictive insights guide the scheduling model to maximize impressions while minimizing post overlaps and balancing visibility across the network. • Real-Time Adaptation: The model incorporates real-time impressions data to refine predictions, enabling the system to dynamically adjust to fluctuations in engagement trends. • Future Expansion: Building on this model, the goal is to integrate a learning system based on the linUCB algorithm, balancing exploration and exploitation to allow the model to better adapt to the evolving dynamics of social media.
Driven by carefully selected data and machine learning techniques, this framework aims to enhance through-partner social engagement. By empowering IT partners to actively participate in vendor-led social media campaigns while avoiding over-publishing, we seek to optimize engagement strategies. Furthermore, the integration of a learning framework aims to enable the project to autonomously adapt to changes in the vendor-partner network’s dynamics. Achieving this level of adaptability will require the development of methods to handle non-stationary environments, where engagement patterns evolve over time. With these advancements, the project could set a new standard for intelligent, responsive marketing strategies in the IT channel industry. Moving beyond traditional automation, this project envisions a finely tuned system capable of sustaining impact in the rapidly evolving digital landscape.
...
To address this issue, this project introduces an advanced model for the vendor-partner network, with several key contributions: • Predictive Analytics: A robust predictor is developed to capture the relationship between historical impressions data and social post characteristics, aiming to forecast future impressions with improved accuracy. • Optimized Scheduling: Predictive insights guide the scheduling model to maximize impressions while minimizing post overlaps and balancing visibility across the network. • Real-Time Adaptation: The model incorporates real-time impressions data to refine predictions, enabling the system to dynamically adjust to fluctuations in engagement trends. • Future Expansion: Building on this model, the goal is to integrate a learning system based on the linUCB algorithm, balancing exploration and exploitation to allow the model to better adapt to the evolving dynamics of social media.
Driven by carefully selected data and machine learning techniques, this framework aims to enhance through-partner social engagement. By empowering IT partners to actively participate in vendor-led social media campaigns while avoiding over-publishing, we seek to optimize engagement strategies. Furthermore, the integration of a learning framework aims to enable the project to autonomously adapt to changes in the vendor-partner network’s dynamics. Achieving this level of adaptability will require the development of methods to handle non-stationary environments, where engagement patterns evolve over time. With these advancements, the project could set a new standard for intelligent, responsive marketing strategies in the IT channel industry. Moving beyond traditional automation, this project envisions a finely tuned system capable of sustaining impact in the rapidly evolving digital landscape.
...
The rise of social media has significantly influenced digital marketing, especially within the IT Channel Industry. Channext, a key player in automating channel marketing, faces the challenge of achieving optimal through-partner social engagement, which includes maximizing shares, impressions, clicks, likes, and comments on social media posts. However, the current approach is hindered by content overload, leading to scheduling conflicts and reduced engagement.
To address this issue, this project introduces an advanced model for the vendor-partner network, with several key contributions: • Predictive Analytics: A robust predictor is developed to capture the relationship between historical impressions data and social post characteristics, aiming to forecast future impressions with improved accuracy. • Optimized Scheduling: Predictive insights guide the scheduling model to maximize impressions while minimizing post overlaps and balancing visibility across the network. • Real-Time Adaptation: The model incorporates real-time impressions data to refine predictions, enabling the system to dynamically adjust to fluctuations in engagement trends. • Future Expansion: Building on this model, the goal is to integrate a learning system based on the linUCB algorithm, balancing exploration and exploitation to allow the model to better adapt to the evolving dynamics of social media.
Driven by carefully selected data and machine learning techniques, this framework aims to enhance through-partner social engagement. By empowering IT partners to actively participate in vendor-led social media campaigns while avoiding over-publishing, we seek to optimize engagement strategies. Furthermore, the integration of a learning framework aims to enable the project to autonomously adapt to changes in the vendor-partner network’s dynamics. Achieving this level of adaptability will require the development of methods to handle non-stationary environments, where engagement patterns evolve over time. With these advancements, the project could set a new standard for intelligent, responsive marketing strategies in the IT channel industry. Moving beyond traditional automation, this project envisions a finely tuned system capable of sustaining impact in the rapidly evolving digital landscape.
To address this issue, this project introduces an advanced model for the vendor-partner network, with several key contributions: • Predictive Analytics: A robust predictor is developed to capture the relationship between historical impressions data and social post characteristics, aiming to forecast future impressions with improved accuracy. • Optimized Scheduling: Predictive insights guide the scheduling model to maximize impressions while minimizing post overlaps and balancing visibility across the network. • Real-Time Adaptation: The model incorporates real-time impressions data to refine predictions, enabling the system to dynamically adjust to fluctuations in engagement trends. • Future Expansion: Building on this model, the goal is to integrate a learning system based on the linUCB algorithm, balancing exploration and exploitation to allow the model to better adapt to the evolving dynamics of social media.
Driven by carefully selected data and machine learning techniques, this framework aims to enhance through-partner social engagement. By empowering IT partners to actively participate in vendor-led social media campaigns while avoiding over-publishing, we seek to optimize engagement strategies. Furthermore, the integration of a learning framework aims to enable the project to autonomously adapt to changes in the vendor-partner network’s dynamics. Achieving this level of adaptability will require the development of methods to handle non-stationary environments, where engagement patterns evolve over time. With these advancements, the project could set a new standard for intelligent, responsive marketing strategies in the IT channel industry. Moving beyond traditional automation, this project envisions a finely tuned system capable of sustaining impact in the rapidly evolving digital landscape.
This thesis concerns the fundamental problem of learning the behavior of decisionmaking agents using only observations of how they act in different situations. As humans, we do it all the time, and have been doing it since birth: think about how a child learns to speak and walk. However, this thesis does not only focus on imitating. We go a step further and try to learn why an agent does what it does. In other words, what is the agent’s objective, which led them to act in a certain way? This is a much harder question, but also much more rewarding when answered correctly: knowing the agent’s motivation allows us not only to imitate but also to understand or even influence the agent’s behavior. For this purpose, we ask the question: what is the agent optimizing for when making decisions? For instance, imagine a consumer agent that buys a certain set of products given their budget. To model the consumer’s behavior, we interpret their action as buying the products with maximum utility, given a limited budget. Thus, learning how the consumer evaluates each product (that is, the utility of each product for the consumer) would allow us to understand, replicate, and possibly influence their behavior. Mathematically, we model the decision process of the agent as an optimization program, and we use Inverse Optimization (IO) as a tool to “reverse engineer” the agent’s optimization program from observed behavior. This thesis can be divided into two parts. First, we dive into the mathematical formalization of the IO problem. We look at the geometry of thema thematical objects emerging from IO problems, and we discuss what it means to solve the IO problem and different ways to do it, proposing tractable reformulations and efficient algorithms. In the second part of this thesis, we develop a tailored IO methodology to solve IO problems emerging from routing problems. We test the potential of our methodology for modeling human driving behavior on real-world problems using data from the Amazon Last Mile Routing Research Challenge. We achieve excellent results, showcasing the potential of IO to solve real-world problems. Additionally, we also developed Inv Opt, an open-source Python package to solve general IO problems.
...
This thesis concerns the fundamental problem of learning the behavior of decisionmaking agents using only observations of how they act in different situations. As humans, we do it all the time, and have been doing it since birth: think about how a child learns to speak and walk. However, this thesis does not only focus on imitating. We go a step further and try to learn why an agent does what it does. In other words, what is the agent’s objective, which led them to act in a certain way? This is a much harder question, but also much more rewarding when answered correctly: knowing the agent’s motivation allows us not only to imitate but also to understand or even influence the agent’s behavior. For this purpose, we ask the question: what is the agent optimizing for when making decisions? For instance, imagine a consumer agent that buys a certain set of products given their budget. To model the consumer’s behavior, we interpret their action as buying the products with maximum utility, given a limited budget. Thus, learning how the consumer evaluates each product (that is, the utility of each product for the consumer) would allow us to understand, replicate, and possibly influence their behavior. Mathematically, we model the decision process of the agent as an optimization program, and we use Inverse Optimization (IO) as a tool to “reverse engineer” the agent’s optimization program from observed behavior. This thesis can be divided into two parts. First, we dive into the mathematical formalization of the IO problem. We look at the geometry of thema thematical objects emerging from IO problems, and we discuss what it means to solve the IO problem and different ways to do it, proposing tractable reformulations and efficient algorithms. In the second part of this thesis, we develop a tailored IO methodology to solve IO problems emerging from routing problems. We test the potential of our methodology for modeling human driving behavior on real-world problems using data from the Amazon Last Mile Routing Research Challenge. We achieve excellent results, showcasing the potential of IO to solve real-world problems. Additionally, we also developed Inv Opt, an open-source Python package to solve general IO problems.
Master thesis
(2023)
-
A. SRIKRISHNA, P. Mohajerin Esfahani, R. Rahimi Baghbadorani, M. Guo, Steven Wilkins
Fuel consumption reduction in Hybrid Electric Vehicles (HEV) powertrains has been an important area of research over the past few decades. HEV powertrains have two energy sources : fuel and battery. The important task of splitting the energy/power demand between both these sources is performed by the Energy Management Systems (EMS). There are many EMS methods and the focus of
this thesis is on a method called Modular ECMS (MEMS) implemented by TNO. MEMS finds the optimal power split among the subsystems by minimizing the energy loss in each subsystem. This strategy assumes that the operating speed of the subsystems of the powertrains is known and uses this knowledge to find the optimal power split and torque among these subsystems. The objective of this thesis is to find the optimal operating speed of the subsystems as well. This is done by a least squares fitting of the objective function and constraints as functions of subsystems speed and torque. A revised Optimal Control Problem (OCP) is formulated as a quadratic programming problem of speed and torque and is termed as Speed-Torque Coupled MEMS (ST-MEMS). The ST-MEMS algorithm is tested on a series-hybrid wheel loader powertrain model and its performance is compared to MEMS, with the model and data provided by TNO. It is concluded that the ST-MEMS, while adding the speed and torque bounds as degrees of freedom, does not achieve a good distribution of power between the 2 sources. The reason for this behaviour is analyzed and an alternate approachis suggested for future work. ...
this thesis is on a method called Modular ECMS (MEMS) implemented by TNO. MEMS finds the optimal power split among the subsystems by minimizing the energy loss in each subsystem. This strategy assumes that the operating speed of the subsystems of the powertrains is known and uses this knowledge to find the optimal power split and torque among these subsystems. The objective of this thesis is to find the optimal operating speed of the subsystems as well. This is done by a least squares fitting of the objective function and constraints as functions of subsystems speed and torque. A revised Optimal Control Problem (OCP) is formulated as a quadratic programming problem of speed and torque and is termed as Speed-Torque Coupled MEMS (ST-MEMS). The ST-MEMS algorithm is tested on a series-hybrid wheel loader powertrain model and its performance is compared to MEMS, with the model and data provided by TNO. It is concluded that the ST-MEMS, while adding the speed and torque bounds as degrees of freedom, does not achieve a good distribution of power between the 2 sources. The reason for this behaviour is analyzed and an alternate approachis suggested for future work. ...
Fuel consumption reduction in Hybrid Electric Vehicles (HEV) powertrains has been an important area of research over the past few decades. HEV powertrains have two energy sources : fuel and battery. The important task of splitting the energy/power demand between both these sources is performed by the Energy Management Systems (EMS). There are many EMS methods and the focus of
this thesis is on a method called Modular ECMS (MEMS) implemented by TNO. MEMS finds the optimal power split among the subsystems by minimizing the energy loss in each subsystem. This strategy assumes that the operating speed of the subsystems of the powertrains is known and uses this knowledge to find the optimal power split and torque among these subsystems. The objective of this thesis is to find the optimal operating speed of the subsystems as well. This is done by a least squares fitting of the objective function and constraints as functions of subsystems speed and torque. A revised Optimal Control Problem (OCP) is formulated as a quadratic programming problem of speed and torque and is termed as Speed-Torque Coupled MEMS (ST-MEMS). The ST-MEMS algorithm is tested on a series-hybrid wheel loader powertrain model and its performance is compared to MEMS, with the model and data provided by TNO. It is concluded that the ST-MEMS, while adding the speed and torque bounds as degrees of freedom, does not achieve a good distribution of power between the 2 sources. The reason for this behaviour is analyzed and an alternate approachis suggested for future work.
this thesis is on a method called Modular ECMS (MEMS) implemented by TNO. MEMS finds the optimal power split among the subsystems by minimizing the energy loss in each subsystem. This strategy assumes that the operating speed of the subsystems of the powertrains is known and uses this knowledge to find the optimal power split and torque among these subsystems. The objective of this thesis is to find the optimal operating speed of the subsystems as well. This is done by a least squares fitting of the objective function and constraints as functions of subsystems speed and torque. A revised Optimal Control Problem (OCP) is formulated as a quadratic programming problem of speed and torque and is termed as Speed-Torque Coupled MEMS (ST-MEMS). The ST-MEMS algorithm is tested on a series-hybrid wheel loader powertrain model and its performance is compared to MEMS, with the model and data provided by TNO. It is concluded that the ST-MEMS, while adding the speed and torque bounds as degrees of freedom, does not achieve a good distribution of power between the 2 sources. The reason for this behaviour is analyzed and an alternate approachis suggested for future work.
HIsarna is a revolutionary step towards production of green steel. Due to the complex nature of the process the control of various facets of HIsarna is difficult. One of these facets is the slag-composition control, or more specifically slag-basicity. In this thesis an optimal sequential decision making strategy has been developed using the Bellman Optimality rule and Value-Iteration. The algorithm has been evaluated at length in order to refine the design choices and decison tree based classification has been used to make the output interpretable for a wider audience. Simulations show the effectiveness of the proposed strategy over the existing control techniques.
...
HIsarna is a revolutionary step towards production of green steel. Due to the complex nature of the process the control of various facets of HIsarna is difficult. One of these facets is the slag-composition control, or more specifically slag-basicity. In this thesis an optimal sequential decision making strategy has been developed using the Bellman Optimality rule and Value-Iteration. The algorithm has been evaluated at length in order to refine the design choices and decison tree based classification has been used to make the output interpretable for a wider audience. Simulations show the effectiveness of the proposed strategy over the existing control techniques.
Monitoring Techniques in Modern Industrial Systems
Fault detection and non-intrusive load monitoring
The Monitoring technique plays a vital role in ensuring the proper functioning of modern industrial systems that are highly sophisticated and automated. On two different applications, this thesis investigates two major categories of information redundancy monitoring techniques, model-based and data-driven.
The first application focuses on ground fault detection in microgrid systems. Leveraging the model information of the system, we propose a design approach for the fault detection filter by creating a linear programming problem. This design ensures the complete decoupling of the disturbance and guarantees fault sensitivity. Recognizing that decoupling is not always feasible, we create a new optimization problem by exploiting available disturbance patterns, so that the filter suppresses the impact of the disturbances while ensuring the fault sensitivity. Simulation studies validate the effectiveness of the proposed designs. The second application deals with non-intrusive load monitoring (NILM) in building systems. Our approach involves a two-stage process that utilizes data to perform NILM. In the first stage, events are identified from the aggregate load measurement. In the second stage, an integer programming problem is formulated to estimate the load for each appliance. The effectiveness of our method is evaluated on a real-world dataset and compared with several other NILM approaches, demonstrating competitive performance in terms of accuracy and computational complexity. ...
The first application focuses on ground fault detection in microgrid systems. Leveraging the model information of the system, we propose a design approach for the fault detection filter by creating a linear programming problem. This design ensures the complete decoupling of the disturbance and guarantees fault sensitivity. Recognizing that decoupling is not always feasible, we create a new optimization problem by exploiting available disturbance patterns, so that the filter suppresses the impact of the disturbances while ensuring the fault sensitivity. Simulation studies validate the effectiveness of the proposed designs. The second application deals with non-intrusive load monitoring (NILM) in building systems. Our approach involves a two-stage process that utilizes data to perform NILM. In the first stage, events are identified from the aggregate load measurement. In the second stage, an integer programming problem is formulated to estimate the load for each appliance. The effectiveness of our method is evaluated on a real-world dataset and compared with several other NILM approaches, demonstrating competitive performance in terms of accuracy and computational complexity. ...
The Monitoring technique plays a vital role in ensuring the proper functioning of modern industrial systems that are highly sophisticated and automated. On two different applications, this thesis investigates two major categories of information redundancy monitoring techniques, model-based and data-driven.
The first application focuses on ground fault detection in microgrid systems. Leveraging the model information of the system, we propose a design approach for the fault detection filter by creating a linear programming problem. This design ensures the complete decoupling of the disturbance and guarantees fault sensitivity. Recognizing that decoupling is not always feasible, we create a new optimization problem by exploiting available disturbance patterns, so that the filter suppresses the impact of the disturbances while ensuring the fault sensitivity. Simulation studies validate the effectiveness of the proposed designs. The second application deals with non-intrusive load monitoring (NILM) in building systems. Our approach involves a two-stage process that utilizes data to perform NILM. In the first stage, events are identified from the aggregate load measurement. In the second stage, an integer programming problem is formulated to estimate the load for each appliance. The effectiveness of our method is evaluated on a real-world dataset and compared with several other NILM approaches, demonstrating competitive performance in terms of accuracy and computational complexity.
The first application focuses on ground fault detection in microgrid systems. Leveraging the model information of the system, we propose a design approach for the fault detection filter by creating a linear programming problem. This design ensures the complete decoupling of the disturbance and guarantees fault sensitivity. Recognizing that decoupling is not always feasible, we create a new optimization problem by exploiting available disturbance patterns, so that the filter suppresses the impact of the disturbances while ensuring the fault sensitivity. Simulation studies validate the effectiveness of the proposed designs. The second application deals with non-intrusive load monitoring (NILM) in building systems. Our approach involves a two-stage process that utilizes data to perform NILM. In the first stage, events are identified from the aggregate load measurement. In the second stage, an integer programming problem is formulated to estimate the load for each appliance. The effectiveness of our method is evaluated on a real-world dataset and compared with several other NILM approaches, demonstrating competitive performance in terms of accuracy and computational complexity.
Large-Scale Setpoint Tracking Controller for Co-regulation of Electric Vehicle Charging Stations
Coordinating Charging with Energy Market Dynamics
This research explores the feasibility of building a large-scale setpoint tracking controller for the co-regulation of Electric Vehicle (EV) charging stations, aiming to coordinate charging with energy market dynamics and minimize the error between a power setpoint and the aggregated consumption of charging stations while capitalizing on developments in the imbalance market. The study examines the roles of actors in the energy market, the characteristics of the EV charging infrastructure, and the information provided by TenneT regarding the imbalance market. Using historical charging data and information provided by TenneT regarding the imbalance market, an optimization problem is formulated and a method for coordinating EV charging is proposed. Our sensitivity analysis of the weight parameters and reduction factor shows their significant impact on the performance of the controller. In this study, we evaluate the performance of the proposed co-regulation controller by tuning the weight parameters to find the optimal balance between financial benefits and customer satisfaction. Our sensitivity analysis of the weight parameters demonstrates that changing them can have a significant impact on the performance of the controller. We also consider the impact of the reduction factor on the performance of the controller and find that increasing it enhances financial benefits but reduces customer satisfaction. Our simulation results indicate that the proposed co-regulation controller can effectively balance financial benefits and customer satisfaction by using appropriate weight parameters. We estimate a yearly profit of e266.45 per EV user, which is equivalent to 13.2% reduction in cost. In conclusion, our research demonstrates the feasibility and effectiveness of using co-regulation to manage the charging demand of electric vehicles in a cost-effective and sustainable way. Our findings provide valuable insight for the development of smart charging strategies that balance the needs of the EV driver, the grid, and other stakeholders, and have important implications for the energy market. Further research is needed to evaluate the effectiveness and robustness of the proposed solution under varying degrees of uncertainty in the input data. Our proposed solution provides a practical and scalable method for managing the charging demand of electric vehicles and has the potential to contribute significantly to global efforts to reduce carbon emissions.
...
This research explores the feasibility of building a large-scale setpoint tracking controller for the co-regulation of Electric Vehicle (EV) charging stations, aiming to coordinate charging with energy market dynamics and minimize the error between a power setpoint and the aggregated consumption of charging stations while capitalizing on developments in the imbalance market. The study examines the roles of actors in the energy market, the characteristics of the EV charging infrastructure, and the information provided by TenneT regarding the imbalance market. Using historical charging data and information provided by TenneT regarding the imbalance market, an optimization problem is formulated and a method for coordinating EV charging is proposed. Our sensitivity analysis of the weight parameters and reduction factor shows their significant impact on the performance of the controller. In this study, we evaluate the performance of the proposed co-regulation controller by tuning the weight parameters to find the optimal balance between financial benefits and customer satisfaction. Our sensitivity analysis of the weight parameters demonstrates that changing them can have a significant impact on the performance of the controller. We also consider the impact of the reduction factor on the performance of the controller and find that increasing it enhances financial benefits but reduces customer satisfaction. Our simulation results indicate that the proposed co-regulation controller can effectively balance financial benefits and customer satisfaction by using appropriate weight parameters. We estimate a yearly profit of e266.45 per EV user, which is equivalent to 13.2% reduction in cost. In conclusion, our research demonstrates the feasibility and effectiveness of using co-regulation to manage the charging demand of electric vehicles in a cost-effective and sustainable way. Our findings provide valuable insight for the development of smart charging strategies that balance the needs of the EV driver, the grid, and other stakeholders, and have important implications for the energy market. Further research is needed to evaluate the effectiveness and robustness of the proposed solution under varying degrees of uncertainty in the input data. Our proposed solution provides a practical and scalable method for managing the charging demand of electric vehicles and has the potential to contribute significantly to global efforts to reduce carbon emissions.
On the road from Model-Based Dynamical Programming to Model-Free Reinforcement Learning
A sample efficient approach
This thesis introduces a new method, called Mixed Iteration, for controlling Markov Decision Processes when partial information is known about the dynamics of the Markov Decision Process. The algorithm uses sampling to calculate the expectation of partially known dynamics in stochastic environments. Its goal is to lower the number of iterations and computational steps required for convergence compared to traditional model-free algorithms. By lowering the number of samples required to achieve convergence Markov Decision Processes can be controlled and trained more efficiently. Additionally, the thesis discusses how this algorithm can enhance the sample efficiency and convergence rate of Reinforcement Learning algorithms like Q-Learning. The effectiveness of the proposed method will be evaluated in standard Reinforcement Learning problems and compared with the performance of Q-learning. The results show that under certain conditions that will be discussed in the thesis, the new proposed algorithm outperforms classical algorithms in terms of sample efficiency. The study will provide insight into the field of previous partial information in Reinforcement Learning alternatives, as well as the challenges that researchers in this field continue to face.
...
This thesis introduces a new method, called Mixed Iteration, for controlling Markov Decision Processes when partial information is known about the dynamics of the Markov Decision Process. The algorithm uses sampling to calculate the expectation of partially known dynamics in stochastic environments. Its goal is to lower the number of iterations and computational steps required for convergence compared to traditional model-free algorithms. By lowering the number of samples required to achieve convergence Markov Decision Processes can be controlled and trained more efficiently. Additionally, the thesis discusses how this algorithm can enhance the sample efficiency and convergence rate of Reinforcement Learning algorithms like Q-Learning. The effectiveness of the proposed method will be evaluated in standard Reinforcement Learning problems and compared with the performance of Q-learning. The results show that under certain conditions that will be discussed in the thesis, the new proposed algorithm outperforms classical algorithms in terms of sample efficiency. The study will provide insight into the field of previous partial information in Reinforcement Learning alternatives, as well as the challenges that researchers in this field continue to face.
This project develops a channel estimation technique for millimeter wave (mmWave) communication systems. Our method exploits the sparse structure in mmWave channels for low training overhead and accounts for the phase errors in the channel measurements due to phase noise at the oscillator. Specifically, in IEEE 802.11ad/ay-based mmWave systems, the phase errors within a beam refinement protocol packet are almost the same, while the errors across different packets are substantially different.
We show that standard compressed sensing algorithms that treat phase noise as a constant fail when channel measurements are acquired over multiple beam refinement protocol packets. Most of the methods that have addressed this problem treat phase noise as purely random, missing the inherent structure within the measurement packets. We present a novel algorithm called partially coherent matching pursuit for sparse channel estimation under practical phase noise perturbations. The proposed approach leverages this partially coherent structure in the phase errors across multiple packets. Our algorithm iteratively detects the support of sparse signal and employs alternating minimization to jointly estimate the signal and the phase errors.
We numerically show that our algorithm can reconstruct the channel accurately at a lower complexity than the benchmarks, and derive a preliminary support detection bound as a performance guarantee. ...
We show that standard compressed sensing algorithms that treat phase noise as a constant fail when channel measurements are acquired over multiple beam refinement protocol packets. Most of the methods that have addressed this problem treat phase noise as purely random, missing the inherent structure within the measurement packets. We present a novel algorithm called partially coherent matching pursuit for sparse channel estimation under practical phase noise perturbations. The proposed approach leverages this partially coherent structure in the phase errors across multiple packets. Our algorithm iteratively detects the support of sparse signal and employs alternating minimization to jointly estimate the signal and the phase errors.
We numerically show that our algorithm can reconstruct the channel accurately at a lower complexity than the benchmarks, and derive a preliminary support detection bound as a performance guarantee. ...
This project develops a channel estimation technique for millimeter wave (mmWave) communication systems. Our method exploits the sparse structure in mmWave channels for low training overhead and accounts for the phase errors in the channel measurements due to phase noise at the oscillator. Specifically, in IEEE 802.11ad/ay-based mmWave systems, the phase errors within a beam refinement protocol packet are almost the same, while the errors across different packets are substantially different.
We show that standard compressed sensing algorithms that treat phase noise as a constant fail when channel measurements are acquired over multiple beam refinement protocol packets. Most of the methods that have addressed this problem treat phase noise as purely random, missing the inherent structure within the measurement packets. We present a novel algorithm called partially coherent matching pursuit for sparse channel estimation under practical phase noise perturbations. The proposed approach leverages this partially coherent structure in the phase errors across multiple packets. Our algorithm iteratively detects the support of sparse signal and employs alternating minimization to jointly estimate the signal and the phase errors.
We numerically show that our algorithm can reconstruct the channel accurately at a lower complexity than the benchmarks, and derive a preliminary support detection bound as a performance guarantee.
We show that standard compressed sensing algorithms that treat phase noise as a constant fail when channel measurements are acquired over multiple beam refinement protocol packets. Most of the methods that have addressed this problem treat phase noise as purely random, missing the inherent structure within the measurement packets. We present a novel algorithm called partially coherent matching pursuit for sparse channel estimation under practical phase noise perturbations. The proposed approach leverages this partially coherent structure in the phase errors across multiple packets. Our algorithm iteratively detects the support of sparse signal and employs alternating minimization to jointly estimate the signal and the phase errors.
We numerically show that our algorithm can reconstruct the channel accurately at a lower complexity than the benchmarks, and derive a preliminary support detection bound as a performance guarantee.
Optimization-based Approaches for Fault Detection and Estimation
With applications to health-monitoring of energy systems
Advancements in technology and societal demands have led to increasing complexity, size, and automation in modern industrial systems. This trend makes these systems more safety-critical, as the occurrence of faults in system components or subsystems may cause the entire system to fail, resulting in significant economic losses and casualties. Consequently, developing an effective fault diagnosis method is crucial for ensuring the reliability, safety, and performance of industrial systems, especially energy systems, which are so relevant to our lives. However, most model-based fault diagnosis systems developed based on observers and parity space relations have the same order as that of the system. This can cause a significant computational burden when dealing with large-scale and high-dimensional systems. This thesis is dedicated to the design of fault diagnosis filters in the framework of differential-algebraic equations, which produce scalable residual generators with design flexibility. Meanwhile, we consider the impact of disturbances and stochastic noise ondiagnosis results, as well as the fault diagnosis problem within the finite frequency domain. In order to design filters capable of handling these issues, we solve filter parameters through optimization problems that are constructed based on specific diagnosis requirements.
...
Advancements in technology and societal demands have led to increasing complexity, size, and automation in modern industrial systems. This trend makes these systems more safety-critical, as the occurrence of faults in system components or subsystems may cause the entire system to fail, resulting in significant economic losses and casualties. Consequently, developing an effective fault diagnosis method is crucial for ensuring the reliability, safety, and performance of industrial systems, especially energy systems, which are so relevant to our lives. However, most model-based fault diagnosis systems developed based on observers and parity space relations have the same order as that of the system. This can cause a significant computational burden when dealing with large-scale and high-dimensional systems. This thesis is dedicated to the design of fault diagnosis filters in the framework of differential-algebraic equations, which produce scalable residual generators with design flexibility. Meanwhile, we consider the impact of disturbances and stochastic noise ondiagnosis results, as well as the fault diagnosis problem within the finite frequency domain. In order to design filters capable of handling these issues, we solve filter parameters through optimization problems that are constructed based on specific diagnosis requirements.
Optimal bidirectional charging control of Electric Vehicles
Minimizing carbon footprint in a realistic simulation environment
Master thesis
(2022)
-
B.A. Swens, S. Grammatico, P. Mohajerin Esfahani, Simon H. Tindemans, H. Keemink
Electricity grids worldwide are experiencing increased peak demands and decreasing simultaneity due to higher shares of Renewable Energy Sources (RES). It is expected that many grids will soon reach their limits. One solution to mitigate these issues is exploiting flexibility in e.g. electric vehicles. In this work, a shrinking horizon model predictive controller is constructed to optimally charge and discharge EVs with respect to the day ahead electricity price or grid carbon intensity. The model takes into account that users with a dual rate electricity plan only want to charge during their off-peak hours. A feature to implement a household PV setup in the optimization is included. The possible consequences in terms of associated carbon emissions, utility costs and user costs are analysed using a simulation based on data from 4279 charging sessions that took place between June 23, 2021 and June 23, 2022. The sessions are split in 2855 weekday sessions (duration between 4 and 24 hours) and 1424 weekend sessions (duration between 4 and 60 hours). It is found that using current circumstances, minimizing the carbon emissions using bidirectional charging results in a higher price (5.5 %) for the utility than using the current state of the art, unidirectional charging minimizing the wholesale electricity cost. Bidirectional charging minimizing the wholesale electricity cost results in higher emissions compared to unidirectional charging (2.8 %), and even compared to uncontrolled charging (0.9 – 3.6 %). The reason for this seems to be a negative correlation between carbon intensity and wholesale price during the times that vehicles are typically connected although this needs further investigation to be confirmed.
...
Electricity grids worldwide are experiencing increased peak demands and decreasing simultaneity due to higher shares of Renewable Energy Sources (RES). It is expected that many grids will soon reach their limits. One solution to mitigate these issues is exploiting flexibility in e.g. electric vehicles. In this work, a shrinking horizon model predictive controller is constructed to optimally charge and discharge EVs with respect to the day ahead electricity price or grid carbon intensity. The model takes into account that users with a dual rate electricity plan only want to charge during their off-peak hours. A feature to implement a household PV setup in the optimization is included. The possible consequences in terms of associated carbon emissions, utility costs and user costs are analysed using a simulation based on data from 4279 charging sessions that took place between June 23, 2021 and June 23, 2022. The sessions are split in 2855 weekday sessions (duration between 4 and 24 hours) and 1424 weekend sessions (duration between 4 and 60 hours). It is found that using current circumstances, minimizing the carbon emissions using bidirectional charging results in a higher price (5.5 %) for the utility than using the current state of the art, unidirectional charging minimizing the wholesale electricity cost. Bidirectional charging minimizing the wholesale electricity cost results in higher emissions compared to unidirectional charging (2.8 %), and even compared to uncontrolled charging (0.9 – 3.6 %). The reason for this seems to be a negative correlation between carbon intensity and wholesale price during the times that vehicles are typically connected although this needs further investigation to be confirmed.
Cycling is an increasingly attractive transportation mode, thanks to its health and environmental benefits. Personalized travel assistance services can help make cycling more appealing by providing speed or route advices that can reduce travel time and increase safety while taking into account the personal preferences of cyclists. Due to its ability to learn agents' reward function, Inverse Reinforcement Learning is a suitable algorithm for learning cycling preferences from data.
This thesis aims to describe cycling styles as a set of cycling preferences encoded as a reward function composed of a weighted sum of features. The weights associated to the features composing the reward function represent the importance given to each cycling preference and express the trade-off between different goals of a cyclist. Continuous-time Inverse Reinforcement Learning extracts the weights from empirical cyclists' trajectories collected during an experiment performed in Delft. During the experiment, cyclists were asked to cycle according to three different cycling styles: cautious, normal and aggressive. Differences between weight sets extracted for each cycling styles were analyzed by means of the Kruskar-Wallis statistical test and K-Means clustering algorithm, and the averaged weights for each cycling style were used to simulate a set of test trajectories.
It is shown by simulations that the reward function identified for a specific cycling style leads to an improvement in terms of similarity to test trajectories with the same cycling style with respect to the reward functions corresponding to other cycling styles. The statistical analysis shows that the weights of cautious and aggressive cycling styles show statistical differences and define separate clusters. ...
This thesis aims to describe cycling styles as a set of cycling preferences encoded as a reward function composed of a weighted sum of features. The weights associated to the features composing the reward function represent the importance given to each cycling preference and express the trade-off between different goals of a cyclist. Continuous-time Inverse Reinforcement Learning extracts the weights from empirical cyclists' trajectories collected during an experiment performed in Delft. During the experiment, cyclists were asked to cycle according to three different cycling styles: cautious, normal and aggressive. Differences between weight sets extracted for each cycling styles were analyzed by means of the Kruskar-Wallis statistical test and K-Means clustering algorithm, and the averaged weights for each cycling style were used to simulate a set of test trajectories.
It is shown by simulations that the reward function identified for a specific cycling style leads to an improvement in terms of similarity to test trajectories with the same cycling style with respect to the reward functions corresponding to other cycling styles. The statistical analysis shows that the weights of cautious and aggressive cycling styles show statistical differences and define separate clusters. ...
Cycling is an increasingly attractive transportation mode, thanks to its health and environmental benefits. Personalized travel assistance services can help make cycling more appealing by providing speed or route advices that can reduce travel time and increase safety while taking into account the personal preferences of cyclists. Due to its ability to learn agents' reward function, Inverse Reinforcement Learning is a suitable algorithm for learning cycling preferences from data.
This thesis aims to describe cycling styles as a set of cycling preferences encoded as a reward function composed of a weighted sum of features. The weights associated to the features composing the reward function represent the importance given to each cycling preference and express the trade-off between different goals of a cyclist. Continuous-time Inverse Reinforcement Learning extracts the weights from empirical cyclists' trajectories collected during an experiment performed in Delft. During the experiment, cyclists were asked to cycle according to three different cycling styles: cautious, normal and aggressive. Differences between weight sets extracted for each cycling styles were analyzed by means of the Kruskar-Wallis statistical test and K-Means clustering algorithm, and the averaged weights for each cycling style were used to simulate a set of test trajectories.
It is shown by simulations that the reward function identified for a specific cycling style leads to an improvement in terms of similarity to test trajectories with the same cycling style with respect to the reward functions corresponding to other cycling styles. The statistical analysis shows that the weights of cautious and aggressive cycling styles show statistical differences and define separate clusters.
This thesis aims to describe cycling styles as a set of cycling preferences encoded as a reward function composed of a weighted sum of features. The weights associated to the features composing the reward function represent the importance given to each cycling preference and express the trade-off between different goals of a cyclist. Continuous-time Inverse Reinforcement Learning extracts the weights from empirical cyclists' trajectories collected during an experiment performed in Delft. During the experiment, cyclists were asked to cycle according to three different cycling styles: cautious, normal and aggressive. Differences between weight sets extracted for each cycling styles were analyzed by means of the Kruskar-Wallis statistical test and K-Means clustering algorithm, and the averaged weights for each cycling style were used to simulate a set of test trajectories.
It is shown by simulations that the reward function identified for a specific cycling style leads to an improvement in terms of similarity to test trajectories with the same cycling style with respect to the reward functions corresponding to other cycling styles. The statistical analysis shows that the weights of cautious and aggressive cycling styles show statistical differences and define separate clusters.