Which Experiences Are Influential for RL Agents? Efficiently Estimating The Influence of Experiences

2024-05-23

arXiv.org

AI Summary

In reinforcement learning, efficiently evaluating the influence of individual experience replay samples on policy performance remains challenging; conventional leave-one-out (LOO) estimation is computationally infeasible due to its requirement of repeated policy retraining. To address this, we propose Policy Iteration with Turn-over Dropout (PIToD), the first method that integrates policy iteration with experience-level zero-out dropout perturbations. Leveraging gradient sensitivity analysis and Monte Carlo approximation, PIToD enables interpretable, high-efficiency influence estimation without retraining. Empirically, PIToD achieves over 100× speedup versus LOO while preserving strong rank correlation (Spearman ρ > 0.92). When applied to filter out negatively influential transitions, PIToD improves the average performance of suboptimal agents by 37.2% across multiple benchmark tasks, markedly enhancing the quality of replay data utilization.

Technology Category

• Machine Learning: Reinforcement Learning
• Search and Optimization: Learning to Search
• Reasoning under Uncertainty: Stochastic Optimization

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• User Modeling, Personalization and Recommendation: Practical large-scale studies of user experience
• Responsible Web: Human-perceived consequences of algorithmic deployment on the web

Abstract

In reinforcement learning (RL) with experience replay, experiences stored in a replay buffer influence the RL agent's performance. Information about how these experiences influence the agent's performance is valuable for various purposes, such as identifying experiences that negatively influence underperforming agents. One method for estimating the influence of experiences is the leave-one-out (LOO) method. However, this method is usually computationally prohibitive. In this paper, we present Policy Iteration with Turn-over Dropout (PIToD), which efficiently estimates the influence of experiences. We evaluate how accurately PIToD estimates the influence of experiences and its efficiency compared to LOO. We then apply PIToD to amend underperforming RL agents, i.e., we use PIToD to estimate negatively influential experiences for the RL agents and to delete the influence of these experiences. We show that RL agents' performance is significantly improved via amendments with PIToD.

Problem

• Efficiently estimating influence of experiences in RL
• Identifying negatively influential experiences for RL agents
• Improving RL agent performance by amending experiences

Innovation

• Uses Policy Iteration with Turn-over Dropout
• Efficiently estimates experience influence
• Amends agents by deleting negative experiences