MDMD options discovery for accelerating exploration in sparse-reward domains

Salience Interest Option: Temporal abstraction with salience interest functions

Reinforcement Learning (RL) is a significant machine learning subfield that emphasizes learning actions based on environment to obtain optimal behavior policy. RL agents can make decisions at variable time scales in the form of temporal abstractions, also known as options. The issue of discovering options has seen a considerable research effort. Most notably, the Interest Option Critic (IOC) algorithm first extends the initial set to the interest function, providing a method for learning options specialized to certain state space regions. This approach offers a specific attention mechanism for action selection. Unfortunately, this method still suffers from the classic issues of poor data efficiency and lack of flexibility in RL when learning options end-to-end through backpropagation. This paper proposes a new approach called Salience Interest Option Critic (SIOC), which chooses subsets of existing initiation sets for RL. Specifically, these subsets are not learned by backpropagation, which is slow and tends to overfit, but through particle filters. This approach enables the rapid and flexible identification of critical subsets using only reward feedback. We conducted experiments in discrete and continuous domains, and our proposed method demonstrate higher efficiency and flexibility than other methods. The generated options are more valuable within a single task and exhibited greater interpretability and reusability in multi-task learning scenarios.