A Flexible Approach for Designing Optimal Reward Functions

Defining a reward function that, when optimized, results in the rapid acquisition of an optimal policy, is one of the most challenging tasks involved in applying reinforcement learning algorithms. The behavior learned by agents is directly related to the reward function they are using. Existing works on Optimal Reward Problem (ORP) propose mechanisms to design reward functions that facilitate fast learning, but their application is limited to specific sub-classes of single or multi-agent reinforcement learning problems. Moreover, while these methods identify which rewards should be given in which situation, they do not give clues regarding on which features of the state or environment should be used when defining a reward function. In this paper, we address these and other issues of ORP. Experimental results on a gridworld scenario are used to evaluate the efficacy of our approach in designing effective reward functions.