Learning and evolutionary models
In previous chapters we encountered a wide range of types of model, but all share the characteristic that they remain unchanged during the course of the simulation. In this chapter, we consider models that incorporate learning: as the simulation runs parameters change, or even the form of the model itself changes, in response to its environment. These models are based on work in machine learning and optimization, both very active areas of research. This chapter cannot cover all the current approaches and we shall concentrate on two that are influential in current social simulation: the use of artificial neural networks and models based on evolutionary programming.
Both are loosely based on analogies from biology. The brain is composed of cells called neurons, which communicate by means of a dense web of interconnections conveying electrochemical impulses. Each neuron obtains inputs from a number of other neurons, and if it receives an excitatory input of sufficient strength, it ‘fires’ and outputs a pulse to other neurons. The human brain is estimated to contain around 100 million neurons. Learning takes place when two neurons fire at the same time, strengthening the connection between the two and reinforcing that particular pathway. Artificial neural network models are based on a drastic simplification of these biological findings about the brain. Although an artificial neural network typically consists of less than 50 ‘units’, each analogous to a neuron, rather than the 100 million of a human brain, it is capable of learning that when presented with a stimulus it should output an appropriate signal.
The other analogy used to construct learning models is the process of evolution by natural selection. Imagine a large population of rabbits that