An overview of reinforcement learning techniques

Abstract

Writing control code for a system where the optimal solution is not known in advance can be a very time-consuming process. The Artificial Intelligence (AI) methods typically involve designing a set of rules which can be effective in situations where the problem is precisely defined and well understood. As in real world problems the optimal solution is rarely known, the reinforcement learning framework which incorporates trial and error attempts can be used. Reinforcement learning (RL) is a machine learning technique that involves training an agent to make decisions which are based on the feedback it receives from the environment. One important decision to make when designing an RL system is whether to use a single or multiple agents. This decision depends on the type of problem that needs to be solved as well the environment complexity. Having a goal that can be achieved by a single agent (one player) it is recommended to use single-agent RL while if there is a need for coordination between multiple agents (players) then a multi-agent approach is recommended. In this article, the differences between single agent RL and multi agent RL techniques, as well as their advantages and disadvantages have been presented, and insights are provided into when one approach may be more appropriate than the other.