Reinforcement learning (RL) has made remarkable progress in many decision-making tasks, such as Go, game playing, and robotics control. However, classic RL approaches often presume that all actions can be executed an infinite number of times, which is inconsistent with many decision-making scenarios in which actions have limited budgets or execution opportunities. Imagine an agent playing a gunfighting game with limited ammunition. It only fires when the enemy appears in the correct position, making shooting a sparse-executing action. Such sparse-executing action has not been considered by classic RL algorithms in problem formulation or effective algorithms design. This paper attempts to address sparse-executing action issues by first formalizing the problem as a Sparse Action Markov Decision Process (SA-MDP), in which certain actions in the action space can only be executed for limited amounts of time. Then, we propose a policy optimization algorithm called Action Sparsity REgularization (ASRE) that gives each action a distinct preference. ASRE evaluates action sparsity through constrained action sampling and regularizes policy training based on the evaluated action sparsity, represented by action distribution. Experiments on tasks with known sparse-executing actions, where classical RL algorithms struggle to train policy efficiently, ASRE effectively constrains the action sampling and outperforms baselines. Moreover, we present that ASRE can generally improve the performance in Atari games, demonstrating its broad applicability
翻译:强化学习(RL)在许多决策任务中取得了显著进展,如Go、游戏游戏和机器人控制等。然而,典型的RL方法往往假定所有行动都可以执行无限次数,这与许多行动预算或执行机会有限的决策方案不相符。想象一个玩枪战游戏的代理人,使用有限的弹药。只有当敌人处于正确位置时,它才会开火,使射杀少效行动。传统的RL算法在问题拟订或有效算法设计中并没有考虑到这种微小的执行行动。本文试图通过首先将问题正规化为Sprass Action Markov 决策程序(SA-MDP)来解决执行不易执行的行动问题。在这种情况下,行动空间中的某些行动只能在有限的时间里执行。然后,我们提出一个政策优化算法,即“行动分量”(ASRE),它通过限制行动抽样和基于评估行动紧张性的政策培训来评估行动紧张性。在行动分配中,它试图先将问题正规化。在已知的Sprass Astraal Agramal Aclas 中,我们用已知的缩算法改进了常规行动。