Eventual Discounting Temporal Logic Counterfactual Experience Replay
This addresses policy optimization challenges in reinforcement learning for complex tasks specified via LTL, representing an incremental improvement with novel components.
The paper tackled the problem of myopic policy optimization for tasks specified with linear temporal logic (LTL) by developing a value-function proxy with eventual discounting to maximize LTL satisfaction probability and a counterfactual experience replay method for off-policy data generation. Experiments in discrete and continuous spaces confirmed the effectiveness of the replay approach.
Linear temporal logic (LTL) offers a simplified way of specifying tasks for policy optimization that may otherwise be difficult to describe with scalar reward functions. However, the standard RL framework can be too myopic to find maximally LTL satisfying policies. This paper makes two contributions. First, we develop a new value-function based proxy, using a technique we call eventual discounting, under which one can find policies that satisfy the LTL specification with highest achievable probability. Second, we develop a new experience replay method for generating off-policy data from on-policy rollouts via counterfactual reasoning on different ways of satisfying the LTL specification. Our experiments, conducted in both discrete and continuous state-action spaces, confirm the effectiveness of our counterfactual experience replay approach.