How to Solve a Rubik's Cube from Scratch

In my final year university project, I successfully trained an AI using reinforcement learning to solve a Rubik's Cube. I based my project on the algorithm outlined in a specific paper.

The algorithm begins by training a neural network to predict the number of moves required to reach the solved position from a given scrambled position. To achieve this, I employed simple value iteration and generated a training dataset on the fly. The cubes were randomly scrambled with depths ranging from 1 to 40. https://mybkexperience.cafe

Once the training phase was complete, the trained neural network served as a heuristic in an A* search algorithm to solve Rubik's Cubes. However, there were modifications made to the classic A* search algorithm, including the incorporation of depth weighting. This trade-off between optimality and speed enhanced the solving process.

The training process involved dedicating approximately 7 days with the utilization of a single Tesla P100 GPU. While parallel training would have been beneficial, its implementation would have required substantial effort, resulting in its omission. As a result, the possibilities for hyperparameter tuning and network architecture experimentation were somewhat limited.