Process.

Game Playing Agents

We created two different AI agents to play Pac-Man before attempting to explain their actions.

Q learning

The Q Learning agent learns from its own decisions, correct and incorrect, from past games to play the game intelligently.

Perceptron

The Perceptron agent learns from human players and copies the decisions they made.

Explanation methods

Printing weights

Displays the highest-weighted input into the neural net to indicate what the AI is thinking.

Heuristic

Observes what Pac-Man is doing and generates an explanation using weights pre-generated weights we created based on our opinions and the rules of Pac-Man

Q Learning

Q Learning neural network that takes in the same features as the Q Learning agent and outputs explanations using weights that were manually trained by our group.