Project Abstract

AI is increasingly being utilised in everyday activities and technologies. It is very strong in many fields. Games can be an excellent way to create a microcosm of the problems that AIs can be trained / coded on. Tetris is an example of this, as a stochastic, fast-paced and in real-time game, developing a way to play Tetris effectively and computationally efficiently could be revolutionary for similar problems that AIs face in Tetris e.g. self driving cars. The goal of this project was to create 2 AIs to play Tetris and compare many of the possible metrics which need to be considered when comparing an AI. Training time, computational efficiency etc. rather than just how high a score can be achieved. The AIs utilised were a heuristic algorithm, calculating the best possible move from a given ruleset by running through each legal move of the next piece in play, and a genetic algorithm which, overtime, evolved the metrics of the heuristic balancing the punishments and rewards based on how effectively they helped improve the performance of the AI. Overall, the genetic algorithm, although stronger at playing the game than the base heuristic algorithm, required much longer to train and test before it became viable for only a slight improvement. I believe that this helps demonstrate how balancing resources and time can be important, focusing on a slight improvement for a minimal impact is not the most effective use of time when a much less complex, yet still competent solution, is available.

Keywords: AI for Games, Artificial Intelligence, Heuristic and Genetic Algorithms

 Conference Details

Session: A

Location: Sir Stanley Clarke Auditorium at 11:00 13:00

Markers: Eike Neumann, Jens Blanck

Course: BSc Computer Science 3yr FT

Future Plans: I’m undecided