Accessible Connect 4 AI

Our solution came in two major parts: a reinforcement learning model and an accessible UI. Each posed its own challenges.

• DQN Model Training: Initial single-model DQN struggled to surpass basic play. Splitting into two models improved vertical threat blocking but horizontal and diagonal defense lagged. We tweaked rewards, epsilon decay, network depth, and replay buffer design—within our compute limits. One original model attempted to use a dataset to learn good board positions from. This dataset can be found here: Data Set
• Hybrid Approaches: Unfortunately, many of our original approaches failed to get past extremely basic strategy. We attempted other more complex approaches to try to solve this problem. This first was an AlphaZero-style approach which used a Monte Carlo Simulaiton to look into future moves, however, this was too computationally heavy. We also attempted a DQN with hard-coded moves to block threats and make winning moves, while this performed better it seemed against the spirit of machine learning, Finally we used calls to “perfect” solved-game API which is able to evaluate every move in order to give a better reward function. The final model that we have utilizes this API.
• Accessible UI: Designed to accept voice commands for column selection and a “read board” command. Prototyped via local Jupyter/VS Code (Colab’s audio APIs are limited), using prerecorded audio.