arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2405.17618cs.LGcs.AI
收藏

对称强化学习损失用于多样化任务和模型规模的鲁棒学习

Symmetric Reinforcement Learning Loss for Robust Learning on Diverse Tasks and Model Scales

Department of Computer Science and Engineering, The Ohio State University(计算机科学与工程系,俄亥俄州立大学);Department of Computer Science(计算机科学系);Engineering, The Ohio State University(工程系,俄亥俄州立大学)

Ju-Seung Byun, Andrew Perrault

AI总结:

本研究提出对称强化学习损失(基于反向交叉熵),通过SA2C和SPPO在离散与连续动作任务及RLHF大语言模型任务中验证了训练稳定性和性能提升。

摘要:

强化学习(RL)训练由于目标移动和高梯度方差等因素而本质上不稳定。从人类反馈中强化学习(RLHF)和从人工智能反馈中强化学习(RLAIF)可能引入额外困难。不同的偏好会使对齐过程复杂化,并且随着大语言模型生成未见过的输出,训练好的奖励模型中的预测错误可能变得更加严重。为了增强训练鲁棒性,RL已采用监督学习中的技术,如集成和层归一化。在这项工作中,我们通过将监督学习中用于噪声数据的反向交叉熵(RCE)改编为定义对称RL损失,来提高RL训练的稳定性。我们展示了在各种任务和规模上的性能改进。我们在离散动作任务(Atari游戏)和连续动作空间任务(MuJoCo基准和Box2D)上使用对称A2C(SA2C)和对称PPO(SPPO)进行实验,在有无添加噪声的情况下,SPPO在不同超参数下表现尤为突出。此外,我们通过SPPO在RLHF任务(如IMDB积极情感和TL;DR摘要任务)中性能的提升,验证了对称RL损失在用于大语言模型时的益处。

英文摘要:

Reinforcement learning (RL) training is inherently unstable due to factors such as moving targets and high gradient variance. Reinforcement Learning from Human Feedback (RLHF) and Reinforcement Learning from AI Feedback (RLAIF) can introduce additional difficulty. Differing preferences can complicate the alignment process, and prediction errors in a trained reward model can become more severe as the LLM generates unseen outputs. To enhance training robustness, RL has adopted techniques from supervised learning, such as ensembles and layer normalization. In this work, we improve the stability of RL training by adapting the reverse cross entropy (RCE) from supervised learning for noisy data to define a symmetric RL loss. We demonstrate performance improvements across various tasks and scales. We conduct experiments in discrete action tasks (Atari games) and continuous action space tasks (MuJoCo benchmark and Box2D) using Symmetric A2C (SA2C) and Symmetric PPO (SPPO), with and without added noise with especially notable performance in SPPO across different hyperparameters. Furthermore, we validate the benefits of the symmetric RL loss when using SPPO for large language models through improved performance in RLHF tasks, such as IMDB positive sentiment sentiment and TL;DR summarization tasks.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑