Genetic Algorithms are an optimization technique capable of obtaining optimum or close-to-optimal results in many difficult search problems. As such, it was a natural research step to apply them in Reinforcement Learning problems, to try to find the best policy that an agent can take. Throughout decades, Genetic Algorithms and Reinforcement Learning have evolved in parallel: this paper examines this common history, starting with traditional approaches that serve as an alternative to popular Temporal Difference methods (for instance Q-Learning), with interesting capabilities such as large state generalization, aliasing robustness and adaptability to changing environments.
The analysis of older approaches is key to understand the achievements of the next studied part, the recent Deep Neuroevolution model for Deep Reinforcement Learning, one of the state-of-the-art techniques in complex tasks such as playing Atari games, featuring a population of mutating neural networks with millions of parameters, stored in a compact seed-based representation. This algorithm produces faster and more accurate results than DQN, but is unable to outperform Rainbow in most tasks. Modifications in the Genetic Algorithm beyond vanilla design may lead to further improved performance.
| Author | Albert Espín |
| Date | April 2019 |
| License | Creative Commons Attribution, Non-Commercial, Non-Derivative |