2. A Framework for GBML

The aim of the framework presented in this section is to structure the range of GBML systems into more specific categories about which we can make more specific observations than we could about GBML systems as a whole. We present two categorisations. In the first (§2.1), GBML systems are classified by their role in learning; specifically their application to i) sub-problems of machine learning, ii) learning itself, or iii) meta-learning. In the second categorisation (§2.2), GBML systems are classified by their high-level algorithmic approach as either Pittsburgh or Michigan systems. Following this, in §2.3 we briefly review ways in which learning and evolution interact and in §2.4 we consider various models of GBML not covered earlier.

Before proceeding we note that evolution can output a huge range of phenotypes, from scalar values to complex learning agents, and that agents can be more or less plastic (independent of evolution). For example, if evolution outputs a fixed hypothesis, that hypothesis has no plasticity. In contrast, evolution can output a neural net which, when trained with backpropagation, can learn much. (In the latter approach evolution may specify the network structure while backpropagation adapts the network weights.)

Structure of GBML Systems

We can divide any evolutionary (meta)-learning system into the following parts: i) Representation, which consists of the genotype (the learner's genes) and phenotype (the learner itself, built according to its genes). In simple cases the genotype and phenotype may be identical, for example with the simple ternary LCS rules of §3.5.2. In other cases the two are very different and the phenotype may be derived through a complex developmental process (as in nature); see §3.4 on developmental encodings for neural networks. ii) Feedback, which consists of the learner's objective function (e.g. the error function in supervised learning) and the fitness function which guides evolution. iii) The production system, which applies the phenotypes to the learning problem. iv) The evolutionary system, which adapts genes.