SW
S.D.C. Wehner
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>
16 records found
1
Quantum networks provide functionality not found in classical networks and realizing their potential requires nodes to execute quantum applications reliably.
Qoala, an execution environment for hybrid quantum-classical applications, runs several program instances concurrently on a single node, where they contend for a limited pool of qubits.
Because qubits are held with exclusive access and cannot be preempted into classical storage without destroying their state, the conditions for deadlock arise, possibly leaving programs unable to proceed.
In this thesis, we implement and compare the three established approaches to the deadlock problem, detection with recovery, prevention, and avoidance, in Qoala.
Waiting carries costs beyond execution time, because qubits in memory decohere.
A blocked program is less likely to succeed, and terminating a program to free up resources discards entanglement that is slow to generate.
We evaluate the strategies on classical and quantum metrics across workloads, study how they scale across various configuration, and introduce a Deadlock Impact Score whose coefficients tune recovery to prioritize either runtime or qubit decoherence during termination.
We find that, with a network schedule, success probabilities are largely similar across strategies, typically within about one percentage point, while avoidance consistently achieves the lowest makespan.
Without a network schedule, success probabilities vary more, by up to 13 percentage points, and prevention performs best, with a makespan comparable to avoidance.
This ordering was broadly stable across the network configurations we tested, though not without exceptions.
Our work enables concurrent execution of workloads in Qoala that previously had to run sequentially and opens the door to further research on concurrency in quantum network nodes. ...
Qoala, an execution environment for hybrid quantum-classical applications, runs several program instances concurrently on a single node, where they contend for a limited pool of qubits.
Because qubits are held with exclusive access and cannot be preempted into classical storage without destroying their state, the conditions for deadlock arise, possibly leaving programs unable to proceed.
In this thesis, we implement and compare the three established approaches to the deadlock problem, detection with recovery, prevention, and avoidance, in Qoala.
Waiting carries costs beyond execution time, because qubits in memory decohere.
A blocked program is less likely to succeed, and terminating a program to free up resources discards entanglement that is slow to generate.
We evaluate the strategies on classical and quantum metrics across workloads, study how they scale across various configuration, and introduce a Deadlock Impact Score whose coefficients tune recovery to prioritize either runtime or qubit decoherence during termination.
We find that, with a network schedule, success probabilities are largely similar across strategies, typically within about one percentage point, while avoidance consistently achieves the lowest makespan.
Without a network schedule, success probabilities vary more, by up to 13 percentage points, and prevention performs best, with a makespan comparable to avoidance.
This ordering was broadly stable across the network configurations we tested, though not without exceptions.
Our work enables concurrent execution of workloads in Qoala that previously had to run sequentially and opens the door to further research on concurrency in quantum network nodes. ...
Quantum networks provide functionality not found in classical networks and realizing their potential requires nodes to execute quantum applications reliably.
Qoala, an execution environment for hybrid quantum-classical applications, runs several program instances concurrently on a single node, where they contend for a limited pool of qubits.
Because qubits are held with exclusive access and cannot be preempted into classical storage without destroying their state, the conditions for deadlock arise, possibly leaving programs unable to proceed.
In this thesis, we implement and compare the three established approaches to the deadlock problem, detection with recovery, prevention, and avoidance, in Qoala.
Waiting carries costs beyond execution time, because qubits in memory decohere.
A blocked program is less likely to succeed, and terminating a program to free up resources discards entanglement that is slow to generate.
We evaluate the strategies on classical and quantum metrics across workloads, study how they scale across various configuration, and introduce a Deadlock Impact Score whose coefficients tune recovery to prioritize either runtime or qubit decoherence during termination.
We find that, with a network schedule, success probabilities are largely similar across strategies, typically within about one percentage point, while avoidance consistently achieves the lowest makespan.
Without a network schedule, success probabilities vary more, by up to 13 percentage points, and prevention performs best, with a makespan comparable to avoidance.
This ordering was broadly stable across the network configurations we tested, though not without exceptions.
Our work enables concurrent execution of workloads in Qoala that previously had to run sequentially and opens the door to further research on concurrency in quantum network nodes.
Qoala, an execution environment for hybrid quantum-classical applications, runs several program instances concurrently on a single node, where they contend for a limited pool of qubits.
Because qubits are held with exclusive access and cannot be preempted into classical storage without destroying their state, the conditions for deadlock arise, possibly leaving programs unable to proceed.
In this thesis, we implement and compare the three established approaches to the deadlock problem, detection with recovery, prevention, and avoidance, in Qoala.
Waiting carries costs beyond execution time, because qubits in memory decohere.
A blocked program is less likely to succeed, and terminating a program to free up resources discards entanglement that is slow to generate.
We evaluate the strategies on classical and quantum metrics across workloads, study how they scale across various configuration, and introduce a Deadlock Impact Score whose coefficients tune recovery to prioritize either runtime or qubit decoherence during termination.
We find that, with a network schedule, success probabilities are largely similar across strategies, typically within about one percentage point, while avoidance consistently achieves the lowest makespan.
Without a network schedule, success probabilities vary more, by up to 13 percentage points, and prevention performs best, with a makespan comparable to avoidance.
This ordering was broadly stable across the network configurations we tested, though not without exceptions.
Our work enables concurrent execution of workloads in Qoala that previously had to run sequentially and opens the door to further research on concurrency in quantum network nodes.
Sparse-Exploration Reinforcement Learning for Control of Quantum Error Correction
Shrinking the exploration gap by perturbing fewer parameters, while still tracking drift
Quantum error correction (QEC) needs a quantum computer's control parameters to stay calibrated as a program runs, but these parameters drift on minute-to-hour timescales, requiring periodic recalibration. Sivak et al. propose a reinforcement-learning agent that re-tunes them online, using QEC detection events as its reward and avoiding pauses for recalibration. We independently replicate their simulation and policy-gradient (PGPE) agent, and propose sparse exploring: rather than perturbing every control parameter in each sample, the agent perturbs only a random subset, lowering the mean error rate while it learns. The price of learning, referred to as the exploration gap, grows with the code distance (the size of the error-correcting code) and as the irreducible error rate (the error floor of perfectly tuned hardware) falls, and it compounds over program length. It is therefore expected to matter more on future hardware, and sparse exploring shrinks it. In simulation, under both idealized sinusoidal and hardware-inspired band-limited 1/f drift, sparse exploring closes roughly 74-85% of the exploration gap, reaching up to 90% for slow drift, independent of code distance (tested up to d=9) and irreducible error rate. We give both a fixed-k variant and an adaptive estimator that picks the sparsity online; the adaptive variant tracks steady drift well but underperforms under sudden fast drift. The result is a cheap addition to online steering that lowers error rates over the multi-day programs useful algorithms require, pending confirmation on real hardware.
...
Quantum error correction (QEC) needs a quantum computer's control parameters to stay calibrated as a program runs, but these parameters drift on minute-to-hour timescales, requiring periodic recalibration. Sivak et al. propose a reinforcement-learning agent that re-tunes them online, using QEC detection events as its reward and avoiding pauses for recalibration. We independently replicate their simulation and policy-gradient (PGPE) agent, and propose sparse exploring: rather than perturbing every control parameter in each sample, the agent perturbs only a random subset, lowering the mean error rate while it learns. The price of learning, referred to as the exploration gap, grows with the code distance (the size of the error-correcting code) and as the irreducible error rate (the error floor of perfectly tuned hardware) falls, and it compounds over program length. It is therefore expected to matter more on future hardware, and sparse exploring shrinks it. In simulation, under both idealized sinusoidal and hardware-inspired band-limited 1/f drift, sparse exploring closes roughly 74-85% of the exploration gap, reaching up to 90% for slow drift, independent of code distance (tested up to d=9) and irreducible error rate. We give both a fixed-k variant and an adaptive estimator that picks the sparsity online; the adaptive variant tracks steady drift well but underperforms under sudden fast drift. The result is a cheap addition to online steering that lowers error rates over the multi-day programs useful algorithms require, pending confirmation on real hardware.
Data-centric AI for QEC
Feature engineering for traditional ML models to solve QEC problem on real data
Quantum error correction (QEC) is one of the core challenges in building scalable quantum computers in the noisy intermediate-scale quantum (NISQ) era. Recently, AI-based QEC decoding has attracted significant interest across industry and academia, yet most efforts focus on large sequential deep learning models trained on simulated data. This leaves a gap in training machine learning decoders on purely real hardware data.
This work presents an empirical evaluation of a prototype data pipeline and hand-crafted feature registry for traditional machine learning-based QEC decoding on real data from QuTech and Google, contributing to a broader model lake vision. Ablation, accuracy decay analysis, and feature importance reveal that final-round and temporal defect features are most critical. Gradient boosting models remain competitive with sequential baselines for short sequences, and learned feature priorities shift with code distance.
Further comparison with LSTM-learned encodings on repetition code data reveals temporal interaction and mixing as the key direction for improving the temporal group in the hand-crafted feature registry. Together, these findings provide interpretable insight into real-data QEC decoding using simple models. ...
This work presents an empirical evaluation of a prototype data pipeline and hand-crafted feature registry for traditional machine learning-based QEC decoding on real data from QuTech and Google, contributing to a broader model lake vision. Ablation, accuracy decay analysis, and feature importance reveal that final-round and temporal defect features are most critical. Gradient boosting models remain competitive with sequential baselines for short sequences, and learned feature priorities shift with code distance.
Further comparison with LSTM-learned encodings on repetition code data reveals temporal interaction and mixing as the key direction for improving the temporal group in the hand-crafted feature registry. Together, these findings provide interpretable insight into real-data QEC decoding using simple models. ...
Quantum error correction (QEC) is one of the core challenges in building scalable quantum computers in the noisy intermediate-scale quantum (NISQ) era. Recently, AI-based QEC decoding has attracted significant interest across industry and academia, yet most efforts focus on large sequential deep learning models trained on simulated data. This leaves a gap in training machine learning decoders on purely real hardware data.
This work presents an empirical evaluation of a prototype data pipeline and hand-crafted feature registry for traditional machine learning-based QEC decoding on real data from QuTech and Google, contributing to a broader model lake vision. Ablation, accuracy decay analysis, and feature importance reveal that final-round and temporal defect features are most critical. Gradient boosting models remain competitive with sequential baselines for short sequences, and learned feature priorities shift with code distance.
Further comparison with LSTM-learned encodings on repetition code data reveals temporal interaction and mixing as the key direction for improving the temporal group in the hand-crafted feature registry. Together, these findings provide interpretable insight into real-data QEC decoding using simple models.
This work presents an empirical evaluation of a prototype data pipeline and hand-crafted feature registry for traditional machine learning-based QEC decoding on real data from QuTech and Google, contributing to a broader model lake vision. Ablation, accuracy decay analysis, and feature importance reveal that final-round and temporal defect features are most critical. Gradient boosting models remain competitive with sequential baselines for short sequences, and learned feature priorities shift with code distance.
Further comparison with LSTM-learned encodings on repetition code data reveals temporal interaction and mixing as the key direction for improving the temporal group in the hand-crafted feature registry. Together, these findings provide interpretable insight into real-data QEC decoding using simple models.
Minimum-weight perfect matching (MWPM), the standard fast
decoder for surface codes, decodes the X and Z syndromes on two independent matching
graphs and so discards the X–Z correlation carried by Y errors. We ask whether a
graph neural network (GNN) that ingests the full detector error model (DEM) can
exploit exactly this discarded structure. On distance-3 rotated surface-code memory
(Stim, circuit-level noise) we train a DEM-informed, DEM-weighted edge-feature GNN
and evaluate the per-round logical error rate εL against two baselines: plain MWPM,
and the stronger belief-matching (correlated MWPM) used by recent learned-decoder
studies. Under unbiased noise the GNN reduces εL by 16.1% versus plain MWPM but
is statistically tied with belief-matching. Under Y-biased noise (bias η = 100 on
the one- and two-qubit channels) the advantage grows sharply,
to 59.3% versus plain MWPM and 22.1% versus belief-matching (3-seed range 22–30%),
and grows further with training data, reaching 67% versus plain MWPM at 12M shots
on a depth-37 round ladder (R2 > 0.99); the advantage itself is depth-stable,
not widening with circuit depth. The gain is thus concentrated where correlated
syndrome structure is largest, precisely the regime matching decoders cannot model,
and we show it is data-bottlenecked rather than architectural: it grows with training
data at a fixed model. We also report a clear boundary: at distance 5 and a fixed
8M-shot training budget the advantage reverses: belief-matching overtakes the GNN,
consistent with data-starvation that worsens with code distance. We therefore frame
the result as an accuracy/architecture study: at d = 3 the GNN learns structure
complementary to matching in the correlated-error regime, with
distance scaling as the central open problem.
...
Minimum-weight perfect matching (MWPM), the standard fast
decoder for surface codes, decodes the X and Z syndromes on two independent matching
graphs and so discards the X–Z correlation carried by Y errors. We ask whether a
graph neural network (GNN) that ingests the full detector error model (DEM) can
exploit exactly this discarded structure. On distance-3 rotated surface-code memory
(Stim, circuit-level noise) we train a DEM-informed, DEM-weighted edge-feature GNN
and evaluate the per-round logical error rate εL against two baselines: plain MWPM,
and the stronger belief-matching (correlated MWPM) used by recent learned-decoder
studies. Under unbiased noise the GNN reduces εL by 16.1% versus plain MWPM but
is statistically tied with belief-matching. Under Y-biased noise (bias η = 100 on
the one- and two-qubit channels) the advantage grows sharply,
to 59.3% versus plain MWPM and 22.1% versus belief-matching (3-seed range 22–30%),
and grows further with training data, reaching 67% versus plain MWPM at 12M shots
on a depth-37 round ladder (R2 > 0.99); the advantage itself is depth-stable,
not widening with circuit depth. The gain is thus concentrated where correlated
syndrome structure is largest, precisely the regime matching decoders cannot model,
and we show it is data-bottlenecked rather than architectural: it grows with training
data at a fixed model. We also report a clear boundary: at distance 5 and a fixed
8M-shot training budget the advantage reverses: belief-matching overtakes the GNN,
consistent with data-starvation that worsens with code distance. We therefore frame
the result as an accuracy/architecture study: at d = 3 the GNN learns structure
complementary to matching in the correlated-error regime, with
distance scaling as the central open problem.
Simulation of quantum circuits with array-like DBMSs
An Empirical Case Study of SciDB versus Relational DBMSs
Quantum circuit simulators running on classical hardware are essential for developing the field while quantum technology matures, but they struggle to scale to the large, memory-intensive states that many circuits produce. Recent work has shown that relational database management systems (RDBMSs) can simulate quantum circuits and that they have an advantage when the workload exceeds the available memory. However, quantum circuits are naturally expressed as tensor contractions, which suggests that array-based database systems - designed to store and operate on multi-dimensional arrays - might be a more natural fit. This possibility has not been systematically evaluated. This paper presents an empirical case study of SciDB, a representative array (tensor-based) DBMS, benchmarked against two relational engines (PostgreSQL and Umbra) on two contrasting circuits: GHZ state preparation, whose states are sparse, and the Quantum Fourier Transform, whose states are dense, across increasing qubit counts. We additionally apply autotuning via MLOS/SMAC to optimise SciDB's configuration. Across all tested cases, SciDB was the slowest engine. The sparse GHZ workload exposes a large fixed per-step overhead, while on the dense QFT, this overhead amortises as the gap to the relational engines narrows from over three orders of magnitude to roughly twofold at 24 qubits against PostgreSQL. Autotuning yielded no improvement over the default configuration, indicating that SciDB's bottleneck lies outside the tuned parameters. We conclude that SciDB offers no advantage over RDBMSs for in-core simulation at these scales, and identify out-of-core simulation - the regime in which database backends are expected to excel - as the central open direction.
...
Quantum circuit simulators running on classical hardware are essential for developing the field while quantum technology matures, but they struggle to scale to the large, memory-intensive states that many circuits produce. Recent work has shown that relational database management systems (RDBMSs) can simulate quantum circuits and that they have an advantage when the workload exceeds the available memory. However, quantum circuits are naturally expressed as tensor contractions, which suggests that array-based database systems - designed to store and operate on multi-dimensional arrays - might be a more natural fit. This possibility has not been systematically evaluated. This paper presents an empirical case study of SciDB, a representative array (tensor-based) DBMS, benchmarked against two relational engines (PostgreSQL and Umbra) on two contrasting circuits: GHZ state preparation, whose states are sparse, and the Quantum Fourier Transform, whose states are dense, across increasing qubit counts. We additionally apply autotuning via MLOS/SMAC to optimise SciDB's configuration. Across all tested cases, SciDB was the slowest engine. The sparse GHZ workload exposes a large fixed per-step overhead, while on the dense QFT, this overhead amortises as the gap to the relational engines narrows from over three orders of magnitude to roughly twofold at 24 qubits against PostgreSQL. Autotuning yielded no improvement over the default configuration, indicating that SciDB's bottleneck lies outside the tuned parameters. We conclude that SciDB offers no advantage over RDBMSs for in-core simulation at these scales, and identify out-of-core simulation - the regime in which database backends are expected to excel - as the central open direction.
Currently explanations for the Not-First/Not-Last propagators for the disjunctive constraint have not been explored thoroughly, and have room for improvement.
In this paper, we look into attempting to give more general explanations by looking through the powerset of tasks and finding a smaller set which still propagates.
We have implemented naive, the state-of-the-art and our subset-finding explanations and compared them all.
Experimentally, around 50\% of the results show a lower amount of conflicts and average literal bound distance, however a majority have a higher runtime.
In addition, the more subsets we consider the bigger the runtime, however it not necessarily decrease in amount of conflicts and average literal bound distance.
...
In this paper, we look into attempting to give more general explanations by looking through the powerset of tasks and finding a smaller set which still propagates.
We have implemented naive, the state-of-the-art and our subset-finding explanations and compared them all.
Experimentally, around 50\% of the results show a lower amount of conflicts and average literal bound distance, however a majority have a higher runtime.
In addition, the more subsets we consider the bigger the runtime, however it not necessarily decrease in amount of conflicts and average literal bound distance.
...
Currently explanations for the Not-First/Not-Last propagators for the disjunctive constraint have not been explored thoroughly, and have room for improvement.
In this paper, we look into attempting to give more general explanations by looking through the powerset of tasks and finding a smaller set which still propagates.
We have implemented naive, the state-of-the-art and our subset-finding explanations and compared them all.
Experimentally, around 50\% of the results show a lower amount of conflicts and average literal bound distance, however a majority have a higher runtime.
In addition, the more subsets we consider the bigger the runtime, however it not necessarily decrease in amount of conflicts and average literal bound distance.
In this paper, we look into attempting to give more general explanations by looking through the powerset of tasks and finding a smaller set which still propagates.
We have implemented naive, the state-of-the-art and our subset-finding explanations and compared them all.
Experimentally, around 50\% of the results show a lower amount of conflicts and average literal bound distance, however a majority have a higher runtime.
In addition, the more subsets we consider the bigger the runtime, however it not necessarily decrease in amount of conflicts and average literal bound distance.
Propagators for Constraint Programming Energetic Reasoning
Exploring the Effect of Explanations for Energetic Reasoning
The cumulative constraint is often used when modeling constraint programming problems, frequently seen in scheduling and planning problems. Energetic reasoning is one of the propagators used to enforce this constraint. However, not much has been done to explore strategies for generating explanations, which are then used by the solver for conflict analysis. This paper addresses this gap by applying strategies used in time-table edge-finding to the energetic reasoning propagator. The strategies are initial bounds relaxations and reducing the overload. Furthermore the paper compares two old strategies (naive and greedy task removal) for reducing the overload and proposes two new ones: greedy task shift and a probabilistic heuristic utilizing the knapsack problem. Results on the MiniZinc RCPSP benchmarks show that the initial bounds adjustments provide great benefit, reducing the number of conflicts by at least twenty-five percent. Reducing the overload provided a small improvement (less than five percent) and results suggest there is not much of a difference between the different strategies.
...
The cumulative constraint is often used when modeling constraint programming problems, frequently seen in scheduling and planning problems. Energetic reasoning is one of the propagators used to enforce this constraint. However, not much has been done to explore strategies for generating explanations, which are then used by the solver for conflict analysis. This paper addresses this gap by applying strategies used in time-table edge-finding to the energetic reasoning propagator. The strategies are initial bounds relaxations and reducing the overload. Furthermore the paper compares two old strategies (naive and greedy task removal) for reducing the overload and proposes two new ones: greedy task shift and a probabilistic heuristic utilizing the knapsack problem. Results on the MiniZinc RCPSP benchmarks show that the initial bounds adjustments provide great benefit, reducing the number of conflicts by at least twenty-five percent. Reducing the overload provided a small improvement (less than five percent) and results suggest there is not much of a difference between the different strategies.
As lazy clause generation has seen much success in recent years, the generation of explanations has become the focus of much research. This paper describes how explanations can be generated for detectable precedences in the disjunctive constraint. We also provide a method to incorporate these explanations into the filtering algorithm proposed by Fahimi et al. [7] by adapting Vilím’s explanations [18]. We proposed two approaches to generating explanations: an approach using only the previously scheduled tasks to explain propagation and an approach using an even smaller subset of tasks combined with explanation lifting. An empirical evaluation of two of our approaches for generating explanations compared to a baseline with naïve explanations found that both approaches performed better in terms of conflicts, LBD, learned clause length and runtime. The most advanced approach of the two (last cluster) performed the best. We believe that using the last cluster approach to generating explanations with other propagators for the disjunctive constraint could be successful.
...
As lazy clause generation has seen much success in recent years, the generation of explanations has become the focus of much research. This paper describes how explanations can be generated for detectable precedences in the disjunctive constraint. We also provide a method to incorporate these explanations into the filtering algorithm proposed by Fahimi et al. [7] by adapting Vilím’s explanations [18]. We proposed two approaches to generating explanations: an approach using only the previously scheduled tasks to explain propagation and an approach using an even smaller subset of tasks combined with explanation lifting. An empirical evaluation of two of our approaches for generating explanations compared to a baseline with naïve explanations found that both approaches performed better in terms of conflicts, LBD, learned clause length and runtime. The most advanced approach of the two (last cluster) performed the best. We believe that using the last cluster approach to generating explanations with other propagators for the disjunctive constraint could be successful.
Explanations have been shown to significantly increase the performance of propagators, when applied to solvers that make use of Lazy Clause Generation. However, to date, there has been little work in exploring explanations for the disjunctive constraint and how they perform compared to simple propagation making use of Naive Explanations. I address this gap by considering an Edge-Finding propagation algorithm for the disjunctive and evaluating the performance of three variants of a solver on job-shop problems, each of them using different explanation techniques, which have been adapted from previous work. I also provide an algorithm serving as an extension to Edge-Finding, which can be used to generate explanations. Then, I compare all approaches in relation to a baseline consisting of a solver that has no support for the disjunctive and uses decomposition. Through experiments, I demonstrate that good explanations provide a significant improvement in terms of recorded metrics compared to the naive method, even if generating them requires additional computational overhead.
...
Explanations have been shown to significantly increase the performance of propagators, when applied to solvers that make use of Lazy Clause Generation. However, to date, there has been little work in exploring explanations for the disjunctive constraint and how they perform compared to simple propagation making use of Naive Explanations. I address this gap by considering an Edge-Finding propagation algorithm for the disjunctive and evaluating the performance of three variants of a solver on job-shop problems, each of them using different explanation techniques, which have been adapted from previous work. I also provide an algorithm serving as an extension to Edge-Finding, which can be used to generate explanations. Then, I compare all approaches in relation to a baseline consisting of a solver that has no support for the disjunctive and uses decomposition. Through experiments, I demonstrate that good explanations provide a significant improvement in terms of recorded metrics compared to the naive method, even if generating them requires additional computational overhead.
Evaluating the Accuracy of User Values Elicited through a Textual Interface
Conducting a user study with a textual interface using questions in isolation to capture user values
This research paper focuses on the accuracy and limitations of user values elicited through a textual interface with questions asked in isolation. The primary objective was to conduct a user study using a textual interface that uses questions in isolation to assess the effectiveness and accuracy of this interface type and questioning style. The study involved exploring various scenarios and associated user values, as well as comparing the textual interface with graphical and audio interfaces tested in four related studies. The user study consisted of 15 participants who interacted with the textual interface. Afterwards, they were tasked with judging and adapting their behaviour models while also evaluating the interface's usability. The findings indicate that while the textual interface demonstrates decent usability, participants did not perceive a strong need for the current system and that compared to other interface types, the textual interface does not yield the most accurate results. This research provides insights into the usability and limitations of a textual interface for eliciting user values. It emphasizes the need for further exploration and development of alternative interface types to enhance accuracy and user engagement.
...
This research paper focuses on the accuracy and limitations of user values elicited through a textual interface with questions asked in isolation. The primary objective was to conduct a user study using a textual interface that uses questions in isolation to assess the effectiveness and accuracy of this interface type and questioning style. The study involved exploring various scenarios and associated user values, as well as comparing the textual interface with graphical and audio interfaces tested in four related studies. The user study consisted of 15 participants who interacted with the textual interface. Afterwards, they were tasked with judging and adapting their behaviour models while also evaluating the interface's usability. The findings indicate that while the textual interface demonstrates decent usability, participants did not perceive a strong need for the current system and that compared to other interface types, the textual interface does not yield the most accurate results. This research provides insights into the usability and limitations of a textual interface for eliciting user values. It emphasizes the need for further exploration and development of alternative interface types to enhance accuracy and user engagement.
The accuracy of an audio interface designed for value elicitation
Eliciting personal values from the users to build responsible AI
Behavior support applications aim to provide personalized and flexible support to users in various domains. To achieve this, understanding users' preferences, values, and context is crucial. Creating user models that incorporate users' norms and values has been proposed as a solution to capture the relationship between desired behaviors and values. However, updating and modifying user models at run-time remains a challenge, as users' norms and values may change over time. This study investigates the accuracy of an audio interface designed to elicit values-related information using isolated questions. This involves designing an audio interface and evaluating its effectiveness through participant interactions, where they are presented with four scenarios. It was found that the audio interface performs above average in terms of usability, as indicated by the System Usability Scale score. The accuracy of the user models is evaluated through the Hamming distance and value differences between the base model and the participant-improved model. Most models required a small number of changes, and when changes were made, they were generally minimal. Additionally, feedback collected through open-ended interview questions lays down a basis for further development. The study contributes to the field by demonstrating the efficacy of the audio interface and its potential for updating user models in real-time. Overall, the research findings support the development of more effective and personalized behavior support applications that can adapt to its users.
...
Behavior support applications aim to provide personalized and flexible support to users in various domains. To achieve this, understanding users' preferences, values, and context is crucial. Creating user models that incorporate users' norms and values has been proposed as a solution to capture the relationship between desired behaviors and values. However, updating and modifying user models at run-time remains a challenge, as users' norms and values may change over time. This study investigates the accuracy of an audio interface designed to elicit values-related information using isolated questions. This involves designing an audio interface and evaluating its effectiveness through participant interactions, where they are presented with four scenarios. It was found that the audio interface performs above average in terms of usability, as indicated by the System Usability Scale score. The accuracy of the user models is evaluated through the Hamming distance and value differences between the base model and the participant-improved model. Most models required a small number of changes, and when changes were made, they were generally minimal. Additionally, feedback collected through open-ended interview questions lays down a basis for further development. The study contributes to the field by demonstrating the efficacy of the audio interface and its potential for updating user models in real-time. Overall, the research findings support the development of more effective and personalized behavior support applications that can adapt to its users.
Individuals seeking a healthier lifestyle can benefit from behavior support agents. Customization and transparency are crucial for system effectiveness. This paper proposes using behavior trees as a user model, with a conversational agent extracting necessary information. The conversational interface enhances transparency, allowing users to understand how the system perceives them. Understanding comparative questions is vital to this approach's success. The objective is to investigate modeling personal values accurately using a conversational agent.
Technologically literate participants engaged in iterative dialogue to elicit a personalized user model. Scenarios explored the impact of contextual factors on value alignment. Results revealed decreased accuracy when more values were affected by contextual factors. Comparative questions were less effective than isolated questioning. System usability was rated poor but approaching acceptability. Larger sample sizes are needed for more comprehensive conclusions.
This research lays the foundation for conversational agents that model personal values within behavior trees, advancing behavior support systems. ...
Technologically literate participants engaged in iterative dialogue to elicit a personalized user model. Scenarios explored the impact of contextual factors on value alignment. Results revealed decreased accuracy when more values were affected by contextual factors. Comparative questions were less effective than isolated questioning. System usability was rated poor but approaching acceptability. Larger sample sizes are needed for more comprehensive conclusions.
This research lays the foundation for conversational agents that model personal values within behavior trees, advancing behavior support systems. ...
Individuals seeking a healthier lifestyle can benefit from behavior support agents. Customization and transparency are crucial for system effectiveness. This paper proposes using behavior trees as a user model, with a conversational agent extracting necessary information. The conversational interface enhances transparency, allowing users to understand how the system perceives them. Understanding comparative questions is vital to this approach's success. The objective is to investigate modeling personal values accurately using a conversational agent.
Technologically literate participants engaged in iterative dialogue to elicit a personalized user model. Scenarios explored the impact of contextual factors on value alignment. Results revealed decreased accuracy when more values were affected by contextual factors. Comparative questions were less effective than isolated questioning. System usability was rated poor but approaching acceptability. Larger sample sizes are needed for more comprehensive conclusions.
This research lays the foundation for conversational agents that model personal values within behavior trees, advancing behavior support systems.
Technologically literate participants engaged in iterative dialogue to elicit a personalized user model. Scenarios explored the impact of contextual factors on value alignment. Results revealed decreased accuracy when more values were affected by contextual factors. Comparative questions were less effective than isolated questioning. System usability was rated poor but approaching acceptability. Larger sample sizes are needed for more comprehensive conclusions.
This research lays the foundation for conversational agents that model personal values within behavior trees, advancing behavior support systems.
The alignment of behavior support systems with our personal values becomes increasingly important as behavior support systems continue to influence our daily lives. The purpose of this paper was to explore the use of graphical interfaces and isolation questions to elicit personal values and build accurate user models. An experiment was conducted in two phases to assess the accuracy and usability of the created interface. A comparison was also made with other types of interfaces developed within the research group. This experiment provided valuable insights but also held some limitations. The findings form a valuable contribution for future research and development in building responsible AI and personalized assistance from behavior support agents.
...
...
The alignment of behavior support systems with our personal values becomes increasingly important as behavior support systems continue to influence our daily lives. The purpose of this paper was to explore the use of graphical interfaces and isolation questions to elicit personal values and build accurate user models. An experiment was conducted in two phases to assess the accuracy and usability of the created interface. A comparison was also made with other types of interfaces developed within the research group. This experiment provided valuable insights but also held some limitations. The findings form a valuable contribution for future research and development in building responsible AI and personalized assistance from behavior support agents.
There has been a lot of research focused on the next generation of the internet, the so-called quantum networks. This analysis has been so far limited to mostly symmetrical architectures, but any near-term realisations of quantum networks using existing fibre topologies will contain asymmetry. In this thesis, we investigate how midpoint asymmetry affects quantum repeater protocols implemented with atomic ensembles. We extend the existing simulation framework to allow for midpoint asymmetry. By simulating asymmetry in elementary links, we show that the performance of an elementary link executing quantum key distribution decreases with an increasing degree of asymmetry. This effect can be mitigated by individual optimisation of photon sources at both ends of the elementary link. We present a way how to reduce the search space of such optimisations by developing a heuristic. The contributions of this thesis provide a crucial starting point for investigations of asymmetry in quantum repeater chains.
...
There has been a lot of research focused on the next generation of the internet, the so-called quantum networks. This analysis has been so far limited to mostly symmetrical architectures, but any near-term realisations of quantum networks using existing fibre topologies will contain asymmetry. In this thesis, we investigate how midpoint asymmetry affects quantum repeater protocols implemented with atomic ensembles. We extend the existing simulation framework to allow for midpoint asymmetry. By simulating asymmetry in elementary links, we show that the performance of an elementary link executing quantum key distribution decreases with an increasing degree of asymmetry. This effect can be mitigated by individual optimisation of photon sources at both ends of the elementary link. We present a way how to reduce the search space of such optimisations by developing a heuristic. The contributions of this thesis provide a crucial starting point for investigations of asymmetry in quantum repeater chains.
Quantum software development is the process of conceiving, specifying, designing, programming, documenting, and testing executable quantum programs that are meant to run on practical quantum hardware. Even though quantum software development research has gained traction over the years, it is still mainly focused on problem analysis, language design, and implementation. Software testing, which is the process of executing a program or application with the intent of finding faults, and verifying that the software product is fit for use is yet to receive substantial attention in quantum software development. This work answers the questions of to what extent are classical software testing techniques transferable to quantum programs, how effectively can they be used, what would an application independent theory for testing quantum software look like, and how practical is it with the current state of physical hardware. Using the principles of spectrum based fault localization, we show that we are able to properly detect and localize bugs in the state teleportation application and blind quantum computing application, even with the presence of noise in the hardware.
...
Quantum software development is the process of conceiving, specifying, designing, programming, documenting, and testing executable quantum programs that are meant to run on practical quantum hardware. Even though quantum software development research has gained traction over the years, it is still mainly focused on problem analysis, language design, and implementation. Software testing, which is the process of executing a program or application with the intent of finding faults, and verifying that the software product is fit for use is yet to receive substantial attention in quantum software development. This work answers the questions of to what extent are classical software testing techniques transferable to quantum programs, how effectively can they be used, what would an application independent theory for testing quantum software look like, and how practical is it with the current state of physical hardware. Using the principles of spectrum based fault localization, we show that we are able to properly detect and localize bugs in the state teleportation application and blind quantum computing application, even with the presence of noise in the hardware.
Preparation of multi-partite entangled quantum states under realistic experimental conditions invariably results in states with non-unit fidelity to the target state. Purification protocols address the need for higher fidelity states than what can be directly prepared. These protocols consume several noisy input states and return an output state of higher fidelity, succeeding probabilistically. We introduce a recurrence based purification protocol for two-colorable graph states on $d$-dimensional quantum systems (qudits). We analyze the performance of the protocol in terms of the minimal required fidelity of input states as well as the expected number of attempts required to successfully reach a specific target fidelity. We find that not only is the purification regime larger for states of greater qudit dimension, but the expected number of attempts to successfully purify a state may be orders of magnitude lower. We develop error thresholds for the protocol with faulty two-qudit operations using a general uncorrelated error model and study the dependence on system dimension and state node number. We observe that the gate error threshold of the protocol improves with increasing dimension and moreover that the threshold depends on the degree of the graph but is otherwise independent of the number of nodes. The qualitative behaviour of the error threshold is captured by an analytically solvable model in which a restricted class of errors is considered. The error thresholds determined here may serve as one benchmark of assessing whether future experimental implementations of two qudit operations function well enough to realize a practical advantage of replacing qubit with qudit states in a multi-partite quantum information protocol.
...
Preparation of multi-partite entangled quantum states under realistic experimental conditions invariably results in states with non-unit fidelity to the target state. Purification protocols address the need for higher fidelity states than what can be directly prepared. These protocols consume several noisy input states and return an output state of higher fidelity, succeeding probabilistically. We introduce a recurrence based purification protocol for two-colorable graph states on $d$-dimensional quantum systems (qudits). We analyze the performance of the protocol in terms of the minimal required fidelity of input states as well as the expected number of attempts required to successfully reach a specific target fidelity. We find that not only is the purification regime larger for states of greater qudit dimension, but the expected number of attempts to successfully purify a state may be orders of magnitude lower. We develop error thresholds for the protocol with faulty two-qudit operations using a general uncorrelated error model and study the dependence on system dimension and state node number. We observe that the gate error threshold of the protocol improves with increasing dimension and moreover that the threshold depends on the degree of the graph but is otherwise independent of the number of nodes. The qualitative behaviour of the error threshold is captured by an analytically solvable model in which a restricted class of errors is considered. The error thresholds determined here may serve as one benchmark of assessing whether future experimental implementations of two qudit operations function well enough to realize a practical advantage of replacing qubit with qudit states in a multi-partite quantum information protocol.