UC-Berkeley: The PacMan Project
Date:
- Implemented various search algorithms, including depth-first, breadth-first, uniform cost, and A*, in the Pacman environment to address navigation and traveling salesman problems. The goal is to obtain optimal paths while strategically avoiding ghosts.
- Encompassed the duality of Pacman as both an adversarial and stochastic search problem. Implemented multiagent minimax and expectimax algorithms, as well as the creation of evaluation functions to improve Pacman’s decision-making abilities in the game.
- Implemented value iteration and Q-learning reinforcement learning algorithms, applying them to the AIMA textbook’s Gridworld, Pacman, and a simulated crawling robot. Also developed an approximate Q-learning agent using feature extractors, achieving successful learning outcomes in various environments.
- Developed classification algorithms using TensorFlow, including gradient descent and stochastic variants for neural network training. Optimized model performance through thoughtful feature design, emphasizing the importance of selecting and extracting relevant features for improved classification accuracy.