(In Proceedings of the 8th International Conference on the Simulation and Synthesis of Living Systems) Stability and task complexity: A neural network model of evolution and learning James R. Watson 1 , Nicholas Geard 1 and Janet Wiles 1,2 1 School of Information Technology and Electrical Engineering 2 School of Psychology The University of Queensland, Australia Abstract Since Hinton and Nowlan introduced the Baldwin ef- fect to the evolutionary computation community, agent- based studies of genetic assimilation have uncovered many details of the dynamic processes involved. In a previous paper, we demonstrated genetic assimilation with a simple food/toxin discrimination task using neu- ral network agents that could evolve their learning rate. The study reported in this paper investigated the ge- netic assimilation of more complex learning tasks. Kauffman’s NK landscape model, which can generate landscapes with a variable degree of correlation, was used to define learning tasks of varying levels of com- plexity. Simulations indicate an increased tendency of genetic assimilation to occur as the complexity of the learning task decreases and the environmental stabil- ity increases. These results are explained in terms of the shifting balance between the evolutionary costs and benefits of learning. Introduction The interaction between evolution and learning has been an area of continual interest to the field of artificial life ever since Hinton and Nowlan’s first computer simu- lation of the Baldwin effect (Hinton & Nowlan 1987). What initially appeared to be a relatively simple pro- cess has been found to be a complex phenomena with many interacting components. The Baldwin effect was first described independently by Morgan (1896), Baldwin (1896) and Osborn (1896), and accounts for the tendency of learned behaviours to become genetically specified without resorting to a Lamarckian justification. Hinton and Nowlan’s intro- duction of this biological concept into the evolutionary computation and artificial life communities marked the beginning of a growing body of research. In Hinton and Nowlan’s original simulation, the bene- fit of being able to learn was that it enabled an individual to solve a ‘needle in a haystack’ task within their life- time. However, the advantage of learning was balanced by a fitness function that favoured individuals who were genetically closer to the solution. This cost of learning was sufficient to encourage the population to move, over time, towards a genetically specified solution. The reason that genetic assimilation occurs at all, given the constraints of Darwinian selection, relies on the balance between the relative benefits and costs asso- ciated with the ability to learn (Mayley 1996). The Bald- win effect can be conceptualized in two distinct phases: (i) initially, the ability to learn a task gives some sub- set of the population a selective advantage, resulting in subsequent generations becoming increasingly dom- inated by individuals with the ability to learn; (ii) once the majority of individuals are able to learn, the costs of learning (e.g., resulting from increased intra- population competition between capable learners) cause selection to favour those individuals who, due to muta- tion and/or recombination, are more genetically predis- posed towards the desired behaviour and therefore don’t have as much to learn. Genetic assimilation occurs when the balance between the benefits and costs of learning shifts so that selective pressures drive the ability to learn out of the population. Empirical demonstrations have shown that genetic as- similation is not always a straightforward process. Hin- ton and Nowlan’s original simulation framework rarely, if ever, results in the complete genetic specification of the solution (see Harvey (1993) for a discussion). Other factors, such as the mutation rate (Fontanari & Meir 1990), selection algorithm (Wiles et al. 2001), the cost of learning (Mayley 1996), the amount of phenotypic plas- ticity (French & Messinger 1994) and population size have been shown to have an effect on the occurrence of genetic assimilation. In a previous paper, Watson and Wiles (2002) demon- strated an equivalent result to that of Hinton and Nowlan (1987) using a population of evolving neural net- works as agents. The task was to learn to discriminate between distinct sets of bit-string representations cor- responding to food and toxin. The task was difficult enough that agents who were able to learn initially had a significant advantage. The learning rate was allowed to evolve, providing an indication of the level of phenotypic plasticity in the population. The practical effect of the learning rate is to amplify an agent’s corrective action to a given situation, therefore the cost of learning arose