Repozytorium publikacji - Politechnika Gdańska

The most popular method for optimizing a certain strategy based on a reward is Reinforcement Learning (RL). Lately, a big challenge for this technique are computer games such as StarCraft II which is a real-time strategy game, created by Blizzard. The main idea of this game is to fight between agents and control objects on the battlefield in order to defeat the enemy. This work concerns creating an autonomous bot using reinforced learning, in particular, the Q-Learning algorithm for playing StarCraft. JamesBot consists of three parts. State Manager processes relevant information from the environment. Decision Manager consists of a table implementation of the Q-Learning algorithm, which assigns actions to states, and the epsilon-greedy strategy, which determines the behavior of the bot. In turn, Action Manager is responsible for executing commands. Testing bots involves fighting the default (simple) agent built into the game. Although JamesBot played better than the default (random) agent, it failed to gain the ability to defeat the opponent. The obtained results, however, are quite promising in terms of the possibilities of further development.

Autorzy

• prof. dr hab. inż. Zdzislaw Kowalczuk link otwiera się w nowej karcie ,
• Jan Cybulski link otwiera się w nowej karcie ,
• dr inż. Michał Czubenko link otwiera się w nowej karcie