Strategic Mastery in Terminal: A Novel Reinforcement Learning Approach to Tower-Defense Gameplay, Shivam Kak, UG '26, Ishaan Javali, UG '26, Yagiz Devre, UG '26 (5EE515B6)
From chess to Atari to AlphaGo, performance in games has been a benchmark of performance for machine learning. Terminal is a tower-defense game where players submit algorithms that play against each other over a diamond-shaped board. The goal of the game is to breach the opponent’s defenses and reach the border of the gameboard. Terminal has hundreds of thousands of annual players, with strategies typically involving thousands of hard-coded, rule-based, logical cases for the tower defense simulation. This is a very limited approach as the number of game board configurations is over 10^250!
Thus far, no Reinforcement Learning-based approaches have been attempted for Terminal. We propose a novel solution to Terminal using methods in Reinforcement Learning (RL) and machine learning that achieves a ranking among the top 300 players world-wide, placing the algorithm in the top 1%.
Moreover, as pioneers in the field, we have curated a publicly-available dataset of 30,000 online matches played by over 1,000 different algorithms that can be used for further research of the Terminal game for RL.
Our approach utilizes a RL technique known as behavioral cloning along with convolutional neural networks, which allows for our AI agent to analyze the game board and learn to mimic expert players from our dataset of online matches. This innovative approach not only demonstrates the effectiveness of RL in Terminal but also contributes to the broader advancement of AI in gaming. Our next steps are to improve the algorithm through thousands of games of self-play.