Evolutionary associative memories through genetic programming

Natural systems apply learning during the process of adaptation, as a way of developing strategies that help to succeed them in highly complex scenarios. In particular, it is said that the plans developed by natural systems are seen as a fundamental aspect in survival. Today, there is a huge interest in attempting to replicate some of their characteristics by imitating the processes of evolution and genetics in artificial systems using the very well-known ideas of evolutionary computing. For example, some models for learning adaptive process are based on the emulation of neural networks that are further evolved by the application of an evolutionary algorithm. In this work, we present the evolution of a kind of neural network that is collectible known as associative memories (AM's) and which are considered as a practical tool for reaching learning tasks in pattern recognition problems. AM's are complex operators, based on simple arithmetical functions, which are used to recall patterns in terms of some input data. AM's are considered as part of artificial neural networks (ANN), mainly due to its primary conception; nevertheless, the idea inherent to their mathematical formulation provides a powerful description that helps to reach a specific goal despite the numerous changes that can happen during its operation. In this chapter, we describe the idea of building new AM's through genetic programming (GP) based on the coevolutionary paradigm. The methodology that is proposed consists in splitting the problem in two populations that are used to evolve simultaneously both processes of association and recall that are commonly used in AM's. Experimental results on binary and real value patterns are provided in order to illustrate the benefits of applying the paradigm of evolutionary computing to the synthesis of associative memories.