arXiv Machine Learning

Revisiting Critic Estimation in Soft Actor-Critic Algorithms

The paper investigates alternative critic designs for the Soft Actor-Critic (SAC) algorithm, exploring learning soft advantages, multi-step return targets, and training multi-step critics from replay buffer data. Six critic variants are compared across fourteen tasks, revealing that training a λ-return critic on replayed trajectory segments yields the most performance gains, while learning soft advantages or using soft GAE offers no benefit. These findings suggest that uncorrected multi-step targets can be effectively applied to SAC and potentially other reinforcement learning algorithms.

arXiv AI
Oct 9

A Geometric Approach to Soft Actor-Critic with Zonotopes for Locomotion Learning

The paper introduces GeZo‑SAC, a variant of Soft Actor‑Critic that incorporates geometric representations (zonotopes) to adapt critic pessimism based on state and action. Each critic predicts a zonotope whose width is subtracted from the critic value as a pessimistic offset, and the disagreement between critics is measured and used to blend the critics via a log‑sum‑exp aggregation. Experiments on MuJoCo‑v5 locomotion tasks show GeZo‑SAC achieving the highest mean return on Ant‑v5 and Hopper‑v5, while also reducing actuator work and action effort and maintaining near‑zero overestimation frequency.

By Panagiotis Roditis, Panagiotis P. Filntisis, Petros Maragos
arXiv AI
Sep 28

PolicyAttention: Softmax Attention Implements Policy Mirror Descent for Closed-Loop Control

The paper investigates whether causal softmax attention can realize policy mirror descent (PMD) as a repeated controller rather than a one‑step algebraic identity. It constructs a fixed causal‑softmax actor–environment–one‑step‑critic protocol, detailing actor, routing, sampling, and normalization residuals, and shows that a frozen one‑step audit model closely approximates PMD. Empirical results demonstrate that the learned actor with an exact one‑step critic achieves median policy loss only about 5% higher than the exact PMD oracle across multiple control settings.

By Yuhe Sui, Yingzhi Tang, Shufang Chen
arXiv Machine Learning
Aug 28

Shared Actors Need Not Share Critics: Effects of Value Mismatch in Parallel Reinforcement Learning

The paper investigates the problem of sharing a single critic across multiple parallel environments in reinforcement learning. It shows that when environments assign different expected returns to the same state, a shared critic must reconcile conflicting value targets, which can distort advantage estimates and misguide policy updates. The authors propose a simple fix—providing the critic with the environment index—demonstrating through bandit models and experiments on CartPole, MuJoCo, BipedalWalker, and 16 Procgen games that this conditional critic stabilizes learning and boosts returns, achieving a 40.8% improvement in aggregate normalized return on unseen levels.

By Zhenya Liu, Yang Meng, Zhuokai Zhao, Xuefeng Liu, Yuxin Chen
arXiv Machine Learning
Oct 12

Cross-Behavior Policy Optimization: Unified Learning from Offline Demonstrations and On- and Off-Policy Rollouts via a Shared Regression Target

The paper introduces Cross-Behavior Policy Optimization (CBPO), a method that unifies offline demonstrations and policy-generated rollouts into a single regression loss using a fixed-reference target. CBPO eliminates the need for separate supervised fine‑tuning and reinforcement learning objectives, allowing positive and negative offline completions, reference responses, and current and historical rollouts to share a common reward‑labeled format and replay buffer. Experiments on two model scales and four mathematical reasoning benchmarks show that CBPO, especially the CBPO‑Unified variant, outperforms baseline methods in avg@32 and pass@32 metrics, with historical rollouts providing additional gains.

By Du Cheng, Junyu Lu, Jingfan Zhang, Yuzhuo Zhang, Haoxuan Jia, Shichao Weng