The Evolutionary Cost of Learning

Traits that are acquired by members of an evolving population during their lifetime, through adaptive processes such as learning, can become genetically speciied in later generations. Thus there is a change in the level of learning in the population over evolutionary time. This paper explores the idea that as well as the beneets to be gained from learning, there may also be costs to be paid for the ability to learn. It is these costs that supply the selection pressure for the genetic assimilation of acquired traits. Two models are presented that attempt to illustrate this assertion. The rst uses Kauuman's NK tness landscapes to show the eeect that both explicit and implicit costs have on the assimilation of learnt traits. A characteristic`hump' is observed in the graph of the level of plasticity in the population showing that learning is rst selected for and then against as evolution progresses. The second model is a practical example in which neural network controllers are evolved for a small mobile robot. Results from this experiment also show the hump. Abstract Traits that are acquired by members of an evolving population during their lifetime, through adap-tive processes such as learning, can become genetically speciied in later generations. Thus there is a change in the level of learning in the population over evolutionary time. This paper explores the idea that as well as the beneets to be gained from learning, there may also be costs to be paid for the ability to learn. It is these costs that supply the selection pressure for the genetic assimilation of acquired traits. Two models are presented that attempt to illustrate this assertion. The rst uses Kauuman's NK tness landscapes to show the effect that both explicit and implicit costs have on the assimilation of learnt traits. A characteris-tic`hump' is observed in the graph of the level of plasticity in the population showing that learning is rst selected for and then against as evolution progresses. The second model is a practical example in which neural network controllers are evolved for a small mobile robot. Results from this experiment also show the hump.