Adaptive-Horizon Model Predictive Control for Modeling Anticipative Behavior in Cyclist Interaction
Y. Huang (TU Delft - Mechanical Engineering)
Christoph M. Konrad – Mentor (TU Delft - Biomechatronics & Human-Machine Control)
J.K. Moore – Mentor (TU Delft - Biomechatronics & Human-Machine Control)
A. Dabiri – Mentor (TU Delft - Team Azita Dabiri)
S.P. Hoogendoorn – Graduation committee member (TU Delft - Traffic Systems Engineering)
More Info
expand_more
Other than for strictly personal use, it is not permitted to download, forward or distribute the text or part of it, without the consent of the author(s) and/or copyright holder(s), unless the work is under an open content license such as Creative Commons.
Abstract
Anticipation is a vital component of cyclist behavior. Empirical evidence consistently suggests that cyclists do not merely react to immediate stimuli, but also actively predict future developments and adjust their motion accordingly. And such anticipative behavior varies across interaction scenarios. However, existing model-based studies provide only limited representations of how cyclists anticipate and how this anticipation varies, often relying on heuristic perception ranges, if-then decision rules, constant preview distance or fixed prediction horizons. Furthermore, anticipation-related parameters are rarely calibrated independently using real traffic data.
To address these gaps, this thesis proposes an adaptive-horizon Model Predictive Control (MPC) framework to explicitly represent cyclists' anticipative behavior. The proposed framework builds upon an existing MPC formulation for cyclist interaction and extends it by introducing an adaptive prediction horizon that varies with the inter-cyclist distance. We formulate the prediction horizon as a bounded and distance-dependent variable that can continuously adapt.
We calibrate the model parameters using two-cyclist overtaking interactions extracted from the TUMDOT dataset. The objective is to optimize MPC parameters to match the measured trajectory of the overtaking cyclist via a two-step Bayesian Optimization (BO) procedure. To disentangle the effects of the horizon-varying parameters from those of the remaining model parameters (background parameters), we implement the calibration sequentially. In step 1, we assume constant-horizon MPC and jointly calibrate the background parameters across all cyclist pairs within a given group. In step 2, the optimal background parameters are fixed, and the horizon-varying parameters are calibrated individually for each cyclist pair. We repeat both calibration steps under 6 BO configurations to reduce sensitivity to BO settings and to mitigate convergence to suboptimal local solutions.
The optimal constant-horizon model calibrated in step 1 achieves an average trajectory error of 0.3691 m relative to the measured trajectories, whereas the varying-horizon model calibrated in step 2 reduces this error to 0.1682 m, corresponding to an improvement of approximately 54.43%. In the correlation analysis of calibrated parameters, the practically used minimum prediction horizon is strongly correlated with its theoretical lower bound, indicating that this bound is often reached in practice. Also, the minimum horizon in all samples occurs at the overtaking point.
Furthermore, we observe moderate correlation between the calibrated horizon-related parameters and certain trajectory features. Higher speed is associated with smaller prediction horizon, indicating a shift toward more reactive behavior when cyclists move faster. And larger minimum inter-cyclist distance is associated with larger prediction horizon, suggesting that stronger anticipation may be associated with maintaining safer spatial margins during overtaking.