Capturing a Recursive Pattern in Neural-Symbolic Reinforcement Learning

Neural-symbolic methods have gained considerable attention in recent years because they are valid approaches to obtain synergistic integration between deep reinforcement learning and symbolic reinforcement learning. Along these lines of research, this paper presents an extension to a recent neural-symbolic method for reinforcement learning. The original method, called State-Driven Neural Logic Reinforcement Learning, generates sets of candidate logic rules from the states of the environment, and it uses a differentiable architecture to select the best subsets of the generated rules that solve the considered training tasks. The proposed extension modifies the rule generation procedure of the original method to effectively capture a recursive pattern among the states of the environment. The experimental results presented in the last part of this paper provide empirical evidence that the proposed approach is beneficial to the learning process. Actually, the proposed extended method is able to tackle diverse tasks while ensuring good generalization capabilities, even in tasks that are problematic for the original method because they exhibit recursive patterns.