Creating Agents to Learn Jenga

Presenter(s)

Pranav Rajbhandari

Abstract or Description

I created AI agents to play the game Jenga autonomously. The game was played in a simulated physics environment. Matches between agents took place in a tournament bracket, with the best adversarial agent taking home trophy.jpg.

I created agents using various methods, including MCTS adversarial search, neural network function approximation (DQNs), and Inverse Reinforcement learning.

I used state estimation to handle the large state space of all possible Jenga tower. I also created an environment model that predicted how likely a tower was to fall over. The agents used this to perform model based reinforcement learning. The environment model was trained on data from a simulator, and used active learning to train alongside the agents actually playing the game.

Mentor

Reid Simmons