MontyBot: An intelligent agent utilizing MCTS and NN to play Starcraft II
Authors
Term
4. term
Education
Publication year
2025
Submitted on
2025-06-05
Pages
58
Abstract
In this project, we sought out to extend our previous work on an economic agent for StarCraft II to an agent capable of playing the full game. We chose to split the game into four distinct subproblems: economic development, army development, combat, and information gathering. The first problem of economic development was handled by the Monte Carlo Tree Search developed in our previous work. To solve the second problem, army development, we decided to extend our Monte Carlo Tree Search. This extension consisted of adding actions to construct production buildings and combat units. Furthermore a representation of the opposing player was added to the state. In this representation, the opponent develops its own army, from which the value of the state will be determined by the probability of winning against the enemy. Our solution to the third subproblem, combat, consisted of two parts. The first is a Combat Prediction Neural Network, used to estimate the probability of winning a combat encounter. The second part handles both offense and defense. It uses the prediction made by the Combat Prediction Neural Network to decide when to attack and how to split up the army to defend. The final subproblem, scouting, was solved by implementing a module to control scout units, and a module to store and infer information. The agent was evaluated by playing games against 9 other agents available on AI Arena. The report concludes by finding that the agent is capable of playing the full game of StarCraft II, with a winrate of 25% on AI Arena.
Keywords
Documents