Institutional Repository of Vilnius University
Autonomous car driving using reinforcement learning and genetic algorithm /
Abstract
dc:descriptionThis study explores applications of three learning strategies for autonomous racing in the TORCS simulation environment - NEAT (NeuroEvolution of Augmenting Topologies), DDPG (Deep Deterministic Policy Gradient), and Evolutionary Reinforcement Learning (ERL) – a hybrid algorithm created during this project. which was created during this project. Each algorithm was trained and tested under controlled conditions to assess performance across three core metrics - reward progression, lap times, and driving behavior. NEAT demonstrated strong early learning and consistent generalization across both seen and unseen tracks but achieved the lowest peak reward and slower lap times. DDPG excelled at fine-tuning policies and produced smoother and faster trajectories but struggled with overfitting and instability. ERL, which integrates NEAT-evolved architectures into DDPG’s gradient-based learning, combined the exploratory strength of NEAT with the policy refinement of DDPG. It outperformed both baselines in peak reward, learning speed, and lap time performance, validating the advantage of hybridizing evolutionary and gradient-based methods. The research confirms the potential of hybrid approaches in autonomous vehicle control and opens pathways for future work focused on multi-phase training strategies.
Degree
thesis:*- Grantor dc:publisher
- Institutional Repository of Vilnius University
- Year dc:date
- 2025
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Milvydas, Žygimantas,
Subjects
dc:subject × 1Rights
dc:rights- Statement dc:rights
-
- info:eu-repo/semantics/openAccess
- Language dc:language
- eng
Identifiers
dc:identifier.*- Repository record dc:identifier
- https://repository.vu.lt/VU:ELABAETD238880340&prefLang=en_US