Automatic Symbolic Modelling of Co-evolutionarily Learned Robot Skills Agapito Ledezma, Antonio Berlanga and Ricardo Aler Universidad Carlos III de Madrid Avda. de la Universidad, 30, 28911, Leganes (Madrid). Spain Abstract Evolutionary based learning systems have proven to be very powerful techniques for solving a wide range of tasks, from prediction to optimization. However, in some cases the learned concepts are unreadable for humans. This prevents a deep semantic analysis of what has been really learned by those systems. We present in this paper an alternative to obtain symbolic models from subsymbolic learning. In the first stage, a subsymbolic learning system is applied to a given task. Then, a symbolic classifier is used for automatically generating the symbolic counterpart of the subsymbolic model. We have tested this approach to obtain a symbolic model of a neural net- work. The neural network defines a simple controller of an autonomous robot. A competitive coevolutive method has been applied in order to learn the right weights of the neural network. The results show that the obtained symbolic model is very accurate in the task of modelling the subsymbolic system, adding to this its readability characteristic. 1 Introduction The use of evolutionary computation (EC) techniques for software development suffers in some aspects from analogous problems to other software development methodologies or paradigms. In particular, we will focus in this paper in the declarative representation of the evolutionary generated descriptions; that is, how we (humans) interpret the output of the EC systems (their generated knowl- edge). In the case of the application we present here, robot control, there are many types of knowledge that could be acquired by means of EC in order to build such systems. Examples are the internal model of robots, models of other robots, communication strategies, or reasoning heuristics. One way of automating this task consists on learning those models by either applying genetic algorithms [1], evolutionary strategies [2], classifier systems [3], or genetic programming [4]. Another view of this type of tasks is centered on the representation structure of the output: the systems can generate rules [5], neural networks [6], etc. When the output is represented in terms of subsymbolic structures (such as neural networks), it is very difficult to interpret the results in order to extract general conclusions on the correctness of the learned knowledge, its possible drawbacks, or the definition of improvements. J. Mira and A. Prieto (Eds.): IWANN 2001, LNCS 2084, pp. 799-806, 200l. © Springer-Verlag Berlin Hdelberg 2001 1