An Adaptive Policy Switching Method for Episodic Cumulative Reward in Reinforcement Learning

In reinforcement learning (RL), agents learn optimal policies through interaction with an environment. Many real-world tasks, such as those in logistics and finance, require achieving cumulative rewards over multiple trials, such as completing work within a specific timeframe or exceeding a target asset value. To address this, we propose Adaptive Policy Switching (APS), a novel method for improving the probability of achieving a target cumulative reward, R, over N trials by adaptively switching between multiple policies. Our method maintains the empirical reward distribution for each policy and selects the policy with the highest probability of achieving the target, considering the remaining trials and cumulative reward. Simulations validate the effectiveness of APS, demonstrating that adaptively switching between policies efficiently achieves cumulative rewards.