FERPO: Forward Entropy-Regularized Policy Optimization
FERPO updates continuous-control policies from critic values without differentiating the critic by actions.
FERPO is an on-policy maximum-entropy reinforcement learning algorithm that derives an entropy- and KL-regularized target action distribution and fits the actor by minimizing a forward-KL objective. The objective is estimated with self-normalized importance sampling from the rollout policy, avoiding differentiation of the critic with respect to actions. Experiments on MuJoCo Playground and ManiSkill show competitive performance and sample-efficiency gains, and actor updates are faster than Relative Entropy Pathwise Policy Optimization.
- Improves the policy from critic values without action derivatives.
- Forward KL covers multiple high-value modes, unlike reverse KL.
- KL limits keep self-normalized importance weights well behaved.
- Competitive sample efficiency on MuJoCo Playground and ManiSkill, with faster actor updates than REPPO.
Full article193 words · extracted from arxiv.org · click to collapse
Several state-of-the-art methods for online reinforcement learning in continuous control improve policies using action gradients of a learned critic. However, critics are typically trained to predict returns, and accurate value predictions do not necessarily yield accurate action derivatives, potentially leading to unreliable policy updates. We propose Forward Entropy-Regularized Policy Optimization (FERPO), an on-policy maximum entropy reinforcement learning algorithm that performs policy improvement using critic values without differentiating the critic with respect to actions. FERPO derives an optimal target action distribution from a policy-improvement objective regularized by entropy and Kullback-Leibler (KL) divergence. We then fit the actor to this target by minimizing a forward-KL objective, estimated using self-normalized importance sampling (SNIS) with actions drawn from the rollout policy. By limiting the target distribution's deviation from the rollout policy, the KL regularization helps keep these importance weights well behaved. In contrast to reverse-KL objectives, which can favor a subset of the target distribution's modes, the forward-KL objective encourages coverage of multiple high-value modes and thereby promotes exploration. Experiments and ablations on MuJoCo Playground and ManiSkill show competitive performance and sample-efficiency gains. Computational benchmarks also demonstrate faster actor updates than Relative Entropy Pathwise Policy Optimization (REPPO).
Text extracted automatically; images, tables and formatting may be missing. Original: https://arxiv.org/abs/2610.02198