A Study on Behavioural Agents for StarCraft 2

Abstract

With the recent trend of artificial intelligence, specifically within machine learning, there are some powerful tools that can be utilized to create video game artificial intelligence bots. Bots that can beat professional players or immerse players within the game to the point where enemies are considered intelligent and react to situations similar to how a real human would. However, some of these processes and tasks to create a bot can be an expensive and time-consuming process. In this research paper, we look at two models to building an AI bot and comparing the two, namely a simple reflex model and a recurrent neural network model. From the results, we can see that the recurrent neural network goes further into the tech tree and is able to produce a more complexed set of units as compared to the simple reflex solution. The simple reflex solution, however, is able to reach the win condition by defeating the enemy bot much quicker than the recurrent neural network solution at 5 min and 39 s and costs less in terms of production and complexity. The recurrent neural network solution was also able to get a higher food supply count and spent the most amount of resources in all areas including technology, economy and army supply.