An Efficient Agent Created In Starcraft 2 Using Pysc2

Abstract

This paper introduces "Judy"(Bot/Agent that is developed using reinforcement learning algorithms using the deep mind toolset PySc2 within StarCraft II. The domain reinforcement learning is a unique uncharted field clustered with difficult problems than worked on in previously done research. By observing the actions and then associating them with reward specs within the SC2 domain we aim to supply an open Python-based GUI which consists of an In-house engine to interact with humans. Adding up to the aforementioned challenges the agent must tackle multiple game maps, using the set of mini-games that specialize in totally different parts of StarCraft II gameplay. The agent relies on game replays of humans who are skilled. By offering a starting or base result for a neural network trained by the acquired info we tend to predict the final outcomes and actions of the players. Thus, the agent experiences a brand new difficulty surrounding the environment exploring deep reinforcement learning algorithms.