Translating the Rashomon Effect to Sequential Decision-Making Tasks

TL;DR

The paper translates the Rashomon effect from classification to sequential decision-making by defining Rashomon sets as collections of policies that, though trained on identical data, induce the same observable behavior (the same induced DTMC) for a given objective but differ in internal representations. It uses probabilistic model checking (via COOL-MC) to verify behavioral equivalence and saliency-based metrics to reveal internal differences, demonstrating the existence of the effect in a taxi domain. Empirically, ensembles drawn from the Rashomon set show robustness under distribution shifts, and a permissive Rashomon policy reduces verification state-space while retaining optimal performance. These findings advance explainable reinforcement learning by showing that multiple, internally diverse explanations can underlie identical policy behavior and that leveraging this diversity can improve verification efficiency and robustness.

Abstract

The Rashomon effect describes the phenomenon where multiple models trained on the same data produce identical predictions while differing in which features they rely on internally. This effect has been studied extensively in classification tasks, but not in sequential decision-making, where an agent learns a policy to achieve an objective by taking actions in an environment. In this paper, we translate the Rashomon effect to sequential decision-making. We define it as multiple policies that exhibit identical behavior, visiting the same states and selecting the same actions, while differing in their internal structure, such as feature attributions. Verifying identical behavior in sequential decision-making differs from classification. In classification, predictions can be directly compared to ground-truth labels. In sequential decision-making with stochastic transitions, the same policy may succeed or fail on any single trajectory due to randomness. We address this using formal verification methods that construct and compare the complete probabilistic behavior of each policy in the environment. Our experiments demonstrate that the Rashomon effect exists in sequential decision-making. We further show that ensembles constructed from the Rashomon set exhibit greater robustness to distribution shifts than individual policies. Additionally, permissive policies derived from the Rashomon set reduce computational requirements for verification while maintaining optimal performance.

Figures (3)

• Figure 1: An example comparison of the Rashomon effect in classification (left) and sequential decision-making (right). In both cases, classification models and COOL-MC with policies produce identical outputs (predictions/induced DTMCs) while differing internally (feature attributions).
• Figure 2: This diagram represents a sequential decision-making system in which an agent interacts with an environment. The agent receives a state and a reward from the environment based on its previous action. The agent then uses this information to select the next action, which it sends to the environment.
• Figure 3: General model checking workflow DBLP:journals/sttt/HenselJKQV22. First, the system needs to be formally modeled, for instance, via PRISM. Then, the requirements are formalized, for instance, via PCTL. Eventually, both are inputted into the model checker, like Storm, which verifies the property.

Theorems & Definitions (5)

• Definition 1: MDP
• Definition 2: DTMC
• Definition 3
• Definition 4: Rashomon Effect in Sequential Decision-Making
• Definition 5: Rashomon Set