Learning Games using a Single Developmental Neuron.
Gul Muhammad Khan, Julian Francis Miller · 2010
An agent controlled by a single developmental neuron is trained to play arcade game. Genetic programming is used to find the DNA of neuron such that it can learn and store the learned information in the form of development in its archi-tecture and updates in chemical concentration. The develop-mental neuron consists of dendrites, axons, and synapses that can grow, change and die. The structure of this neuron com-plexify itself at runtime as a result of game scenarios. The network is tested in arcade game environment of checkers. The agent has to recognize the patterns of the board and use this information to learn how to play the game better. The network is evolved against a professional checker program for its capability to learn. Input from the board is provided using sensory neuron through synapses. The developmen-tal neuron process these signals and send output to the mo-tor neurons to make a move. The structure of the neuron is also modified during signal processing. The developmental neuron successfully defeated the professional minimax based checker program during evolution by a large margin. We also tested the agent against some other opponents (not seen dur-ing evolution) of various levels for its generality and it proves to outperform them.