Reinforcement Learning: Algorithms to balance the bias between Exploration and Exploitation

Reinforcement learning agents map states to optimal actions by developing a policy. In order to learn which of the available actions yields the highest rewards, the agent makes decisions from the available actions. As a result, by choosing the most obvious action to take on each action decision, a Reinforcement Learning agent maximizes its reward. Nevertheless, in order to gain a complete understanding of how favorable each action is, it must explore all the options. Exploration and exploitation are at odds in this area. This bias is balanced by numerous existing algorithms. Epsilon-Greedy Approaches, Optimistically High Initial Value Approaches, and Upper Confidence Bounds Approaches are a few of these techniques. Using the K-Armed Bandit problem as a simulation, this paper compares these algorithms, their benefits and disadvantages. Research suggests that maximizing long-term rewards will require a combination of these methods. Simulations are run to examine the impact of altering the epsilon parameter, and performance changes are noted with decreasing the epsilon parameter within convergence.