Solving NP-Complete Akari games with deep learning

This work proposes a deep learning-based agent that solves 10 x 10 board size puzzle games of the NPComplete Akari genre, while learning tabula rasa, that is, never exposed to any rules of the game. A six-model ensemble is trained to form a single solution policy. Lamp placements are chosen sequentially until the final resolution or failure. For the training and testing of the agent, over 3 million games were randomly produced with at least one known solution. The proposed agent solves 98.1% of Akari games that were never shown to it during training, revealing that the agent can learn implicit rules in NP-complete games. These approaches and training procedures proposed can guide future research on tabula rasa neural network agents for solving games or NP-Complete problems.