Applying Monte Carlo Search and Monte Carlo Tree Search on Embedded Systems to Play Connect Four with a Robotic Arm

Abstract

AbstractMonte Carlo based methods have brought a remarkable improvement in performance of artificial intelligence in the realm of games in recent years, whereby human champions could be beaten in several board games of high complexity. In this work, two Monte Carlo based approaches, the Monte Carlo Search and the Monte Carlo Tree Search, are applied to the game of Connect Four. To test and compare these two approaches, games against each other are simulated. The better approach is then further placed against human opponents using a website (https://carloconnect.com/) and a robotic arm to evaluate the algorithm’s strength. All tests are conducted on a Raspberry Pi 3 Model B to investigate the approaches’ performance in an environment with limited available computing power and memory. The Experiments show that under these conditions the Monte Carlo tree search significantly outperforms the Monte Carlo search and is therefore the better solution. This approach is called CarloConnect and proves itself against human opponents by winning most of the games, thus playing at human level even while having only low computing resources.KeywordsArtificial intelligence for gamesConnect FourMonte Carlo SearchMonte Carlo Tree SearchRaspberry PiRobotic arm