Abstract
This project develops an AI for Versus Tetris using nature-inspired evolutionary algorithms. Tetris is computationally hard (NP-complete), and the adversarial, real-time nature of Versus Tetris introduces additional challenges, including stochastic piece generation and opponent dynamics. We formulate the game AI problem, develop a simulation platform adhering to guideline rules, and implement three optimization algorithms: Genetic Algorithms (GA), Particle Swarm Optimization (PSO), and Ant Colony Optimization (ACO). Each algorithm optimizes weights for a heuristic evaluation function comprising features such as pile height, holes, row transitions, attack potential, and garbage pressure. Experimental evaluation across 30 independent runs measures attack efficiency (APP), survivability, and win-rate against static and dynamic opponents. Results show GA achieves the highest offensive capability (APP: 2.17±0.23), PSO delivers the most balanced performance with superior win-rates (0.73±0.09) and survivability (214.8±27.6 pieces), while ACO exhibits greater variability (APP: 1.68±0.31) but explores novel strategies. Statistical analysis confirms significant differences between algorithms. The findings demonstrate that evolutionary approaches can produce competitive strategies without exhaustive search, contributing insights for adversarial AI domains including cybersecurity and autonomous systems. Future work includes neuro-evolution, co-evolutionary frameworks, and real-time adaptation.