Genetic Action Sequence for Integration of Agent Actions

Abstract

Reinforcement learning has remarkable achievements in areas such as GO, Game, and autonomous vehicles, and is attracting attention as the most promising future technology among neural network-based technique. In spite of its capability, Reinforcement learning suffers from its greedy nature that causes selective behaviors. This is because the agent learns in such a way that each action seeks an optimal reward. However, the pursuit of maximum rewards by each action does not guarantee the maximization of the total reward. This is an obstacle to improving the overall performance of reinforcement learning. This paper introduces a concise illustration of Action Sequence, a method that encourages the trained model to seek grander strategies generating more desirable results. Especially aiming at highly complexed problems where plain actions can hardly handle any tasks. The model was tested under a video game environment to verify its improved proficiency in choosing suitable actions.

This work was supported by GIST Research Institute(GRI) grant funded by the GIST in 2019.