Integrating deep learning and dynamic game theory for trajectory planning

Kurzfassung

This dissertation addresses the challenge of trajectory planning for automated vehicles in multi-agent traffic environments. It presents three papers that explore both game-theoretical approaches and deep learning algorithms, ultimately leading to the development of a trajectory planner known as DeepGame-TP. The primary objective was to create a planner that is interaction-aware, optimal, transparent, interpretable and capable of recognizing and adapting to driver behavior in real time. The first publication introduces a vehicle intention prediction model based on Bayes’ theorem, where the prior is derived from a Mixed Strategy Nash Equilibrium (MSNE) and refined using real time evidence from vehicle maneuvers. Limitations such as a constrained action space composed of pre-computed trajectories and insufficient characterization of driver behavior highlighted the need for further research, leading to the subsequent two papers. The second publication presents an LSTM-based deep learning model for predicting vehicles’ longitudinal behavior. The proposed model predicts the speed profile of each vehicle over a 6 seconds horizon, achieving state of the art performance, particularly within a 4 seconds prediction window, while remaining effective for longer forecasts. Its straightforward input structure allows it to operate using only observable data from the target vehicle, contrasting with many existing algorithms that depend on comprehensive data from surrounding vehicles, an assumption often impractical in real world scenarios. The research concludes in the third paper with the introduction of DeepGame-TP, which integrates the LSTM speed prediction model into a dynamic game framework. DeepGame-TP maintains the transparency of traditional approaches through its dynamic game formulation while benefiting from deep learning to estimate each agent’s cost function. This enables the system to act optimally by recognizing and adapting to the behaviors of different agents. Experiments show its flexibility and reliability across various scenarios and topologies, with computational efficiency that supports real time applications for up to four vehicles. With computational times consistently under 150 milliseconds for these agents, DeepGame-TP represents a robust and adaptive real time solution for autonomous driving, successfully combining game-theoretic transparency with the pattern recognition capabilities of deep learning.