Abstract
This paper presents a novel RL algorithm, S-REINFORCE, designed by leveraging two types of function approximators, namely Neural Network (NN) and Symbolic Regressor (SR), to produce numerical and symbolic policies for dynamic decision-making tasks, respectively. A symbolic policy uncovers functional relations between the underlying states and action-probabilities. Further, the symbolic policy is utilized through importance sampling (IS) to improve the rewards received during the learning process. The effectiveness of S-REINFORCE has been validated on various dynamic decision-making problems involving low and high dimensional action spaces. The results obtained clearly demonstrate that by leveraging the complementary strengths of NN and SR, S-REINFORCE generates policies that exhibit both good performance and interpretability. This makes S-REINFORCE an excellent choice for real-world applications where transparency and causality play a crucial role.