Connect X is a beta-version of a new type of Machine Learning competition called Simulations. In Simulation Competitions, participants compete against a set of rules rather than an evaluation metric. The objective is to create a Python submission file that acts as an "Agent" to play the game Connect X (also known as Connect Four) against a computer or another user.
The Challenge The goal is to get a certain number of your checkers in a row horizontally, vertically, or diagonally on the game board before your opponent. Players take turns dropping one checker into a column at the top of the board. Each move involves strategizing to either win for yourself or prevent your opponent from winning. While the default requirement is four-in-a-row, other options may be introduced later.
How It Differs
Unlike traditional Kaggle competitions where users submit CSV files or Notebooks, this competition requires submitting a Python .py file. The leaderboard is not based on accuracy but on performance against other submissions. Submissions are assigned a Skill Rating modeled by a Gaussian distribution. When uploaded, a submission first plays a validation episode against itself; if it fails, it is marked as an Error. Valid submissions join a pool where they are matched against others with similar ratings. Approximately eight episodes run per day per submission, updating ratings dynamically based on wins, losses, and draws.
Rules for Agents - Submissions must be an "Agent" function that returns the column index to drop a checker. - Agents can only use modules available in the Kaggle Kernels notebook image. - Maximum file size is 100 MB. - An Agent must return an action within 2 seconds (60 seconds on the first turn). - Errors or invalid actions result in a loss for the episode and potential invalidation.
This competition serves as a low-stakes trial run to help the data science community practice and hone skills in Reinforcement Learning.