"What are my options?": Explaining RL Agents with Diverse Near-Optimal Alternatives (Extended)

In this work, we provide an extended discussion of a new approach to explainable Reinforcement Learning called Diverse Near-Optimal Alternatives (DNA), first proposed at L4DC 2025. DNA seeks a set of reasonable "options" for trajectory-planning agents, optimizing policies to produce qualitatively diverse trajectories in Euclidean space. In the spirit of explainability, these distinct policies are used to "explain" an agent's options in terms of available trajectory shapes from which a human user may choose. In particular, DNA applies to value function-based policies on Markov decision processes where agents are limited to continuous trajectories. Here, we describe DNA, which uses reward shaping in local, modified Q-learning problems to solve for distinct policies with guaranteed epsilon-optimality. We show that it successfully returns qualitatively different policies that constitute meaningfully different "options" in simulation, including a brief comparison to related approaches in the stochastic optimization field of Quality Diversity. Beyond the explanatory motivation, this work opens new possibilities for exploration and adaptive planning in RL.

Pointwise-in-Time Explanation for Linear Temporal Logic Rules

This work introduces a framework to assess the relevance of individual linear temporal logic (LTL) constraints at specific times in a given path plan, a task we refer to as "pointwise-in-time" explanation. We develop this framework, featuring a status assessment algorithm, for agents which execute finite plans in a discrete-time, discrete-space setting expressible via a Kripke structure. Given a plan on this structure and a set of LTL rules which are known to constrain the agent, the algorithm responds to two types of user queries to produce explanation. For the selected query time, explanations identify which rules are active, which have just been satisfied, and which are inactive, where the framework status criteria are formally and intuitively defined. Explanations may also include the status of individual rule arguments to provide further insight. In this paper, we systematically present this novel framework and provide an example of its implementation.