Definition of Multi-Objective Deep Reinforcement Learning Reward Functions for Self-Driving Vehicles in the Urban Environment

Abstract

Recent revolutionary advances in cognitive science using the learning principles of biological brains and human cognition have fuelled artificial intelligence (AI), in particular, the development and use of ground-breaking Deep Reinforcement Learning (DRL) in numerous fields by both leveraging the powerful generalisation ability of data-hungry Deep Neural Networks (DNN) and the self-learning ability of Reinforcement Learning (RL). Can DRL provide human-level and beyond better-than-human-level performance in realising the tasks of SDVs by emulating human cognition? This article addresses the aspects of the use of multi-objective (MO) artificial agents (AAs) that can be developed by DRL for managing the real-world dynamics of SDVs in highly dimensional environments - partially observable, multiagent, stochastic, sequential, dynamic, continuous and unknown. Each objective learned separately by the agent is combined to form a broader policy within the proposed techniques. The presented approach is evaluated in urban realistic driving simulations. The simulation results suggest that the proposed DRQN techniques leading to a synergistic policy can perform multiple main driving actions successfully within dense urban traffic scenarios.