Connect4–2-Step Lookahead Moves

This dataset contains 100k actions played by the 2-Step Lookahead Agent in different random 6x7 Connect4 games.

The definition and a basic implementation of this Agent can be found in this Kaggle Tutorial. However, here we implemented a Numpy-version of this Agent, which runs faster and keeps the code simple and efficient. You will find our Numpy implementation (which uses a more advanced strategy) ready for Competition Submission in this notebook (currently 26/204 in the ConnectX Competition Leaderboard, TOP 13%).

This Dataset can be used to train a model (e.g. a Neural Network) to imitate the 2-Step Lookahead Agent. That's a Multi-class Classification Task [Supervised Learning], where our target is the column action (the column chosen by the 2-Step Lookahead Agent). This way, we will have a (faster) model that will know how to do a 2-Step Lookahead search without performing that search actually. If we train a Neural Network, the learned weights will simulate the MinMax Search (if the classification accuracy is high enough).