Skills to Drive: Successor Features for Autonomous Highway Pilot

Reinforcement learning applications are spreading among different domains, including autonomous vehicle control. The diverse situations that can happen during, for instance, at a highway commute infinite, and with labeled data, the perfect coverage of all use-cases sounds ambitious. However, complex tasks complicated scenarios faced during an system design, credit assignment problem arises. How to construct appropriate objectives artificial intelligence learn preferences between goals also matter designer’s choice. This work attempts tackle by utilizing successor features providing possible decomposition reward functions, guiding agent’s actions. method makes training easier agent enables immediate, profound performance on new combined tasks. Furthermore, optimal composition, desired behavior be fine-tuned, as auxiliary gain, empowers driving styles rapidly changeable. We introduce adaptation FastRL algorithm domain, meanwhile developing stabilizing way using Successor Features, namely DoubleFastRL. compare our solution scenario basic agents such Q-learning having multi-objective training.