Learning from a random player using the reference neuron model
George Schleis, Mateen M. Rizki · 2003
The reference neuron model (RNM) is an example of an advanced model of neurological structure. The model is based on the principle of superposition-free memory, i.e., the requirement that the acquisition of new memories by a neural network does not degrade or hybridize the previously acquired memories. The RNM incorporates features found in procedural systems as well as the distributed pattern recognition capabilities found in connectionist neural network approaches. The memory manipulation mechanisms inherent in the model support the development of temporal as well as associative memory structures through trial and error learning. The results of which contribute to the development of knowledge by the model about its environment. The RNM is capable of learning strategies to play board games with a reasonable level of performance. In this work, results of experiments with the game of Tic-tac-toe are presented. The RNM is taught to play the game by simulating contests between the RNM and an artificial opponent that simply makes random moves. Through a series of games, the RNM acquires strategies that permit it to consistently win games. Various sizes of boards were tested to demonstrate that the results produced using RNM scale to larger search spaces. In addition, results are presented that provide evidence that as additional types of memories (e.g. temporal, associative, and artificial imagination) are introduced into the system, the performance improves.