arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2602.01601cs.LGcs.AIcs.CL
收藏

具有可验证奖励的在线强化学习中自适应 rollout 分配

Adaptive Rollout Allocation for Online Reinforcement Learning with Verifiable Rewards

The Chinese University of Hong Kong(香港中文大学);Huawei Hong Kong Research Center(华为香港研究中心)

Hieu Trung Nguyen, Bao Nguyen, Wenao Ma, Yuzhi Zhao, Ruifeng She, Viet Anh Nguyen

AI总结:

VIP通过基于方差的预测分配策略提高在线强化学习中可验证奖励的采样效率和性能。

摘要:

在具有可验证奖励的强化学习中,采样效率是一个关键瓶颈。现有的基于群体的策略优化方法,如GRPO,为所有训练提示分配固定数量的rollouts。这种统一的分配隐式地将所有提示视为等信息量,可能导致计算预算使用效率低下并阻碍训练进度。我们引入了VIP,一种基于方差的预测分配策略,该策略将给定的rollout预算分配给当前批次中的提示,以最小化策略更新的预期梯度方差。在每次迭代中,VIP使用轻量级的高斯过程模型,根据最近的rollouts预测每个提示的成功概率。这些概率预测被转换为方差估计,然后输入一个凸优化问题,以在硬计算预算约束下确定最优的rollout分配。实验证果表明,VIP在多个基准中始终提高了采样效率,并在性能上优于均匀或启发式分配策略。

英文摘要:

Sampling efficiency is a key bottleneck in reinforcement learning with verifiable rewards. Existing group-based policy optimization methods, such as GRPO, allocate a fixed number of rollouts for all training prompts. This uniform allocation implicitly treats all prompts as equally informative, and could lead to inefficient computational budget usage and impede training progress. We introduce VIP, a Variance-Informed Predictive allocation strategy that allocates a given rollout budget to the prompts in the incumbent batch to minimize the expected gradient variance of the policy update. At each iteration, VIP uses a lightweight Gaussian process model to predict per-prompt success probabilities based on recent rollouts. These probability predictions are translated into variance estimates, which are then fed into a convex optimization problem to determine the optimal rollout allocations under a hard compute budget constraint. Empirical results show that VIP consistently improves sampling efficiency and achieves higher performance than uniform or heuristic allocation strategies in multiple benchmarks.

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。

↑