Deep Reinforcement Learning for PlanetWars
The yellow team is powered by a deep neural network that's trained using a monte-carlo style of reinforcement learning. It approximates the value of selecting actions over time, e.g. 7500 games. Network has 350 inputs and 100 outputs, with 3x 1000 hidden layers of rectified linear type. It's trained on the GPU.
There's 50% randomness in the AI, which I only realized later! The first games are lost but the NN wins later anyway despite the randomness...