The fitness of false beliefs

Presents a model in which agents can perform one of four actions, and each action places the agent in a new state. Some states have higher fitness values than others, and the more fit the agent's behaviors are on average, the more likely the agent is to be chosen to reproduce in the next generation. The agent model consists of two parts, both represented by feedforward neural networks. The first part, called the model network, learns to predict which state results from which action. The second part, called the evaluation network, tries to predict the fitness of performing a particular action in a particular state. The evaluation network is not allowed to learn, and so its performance cannot improve during the agent's "lifetime". However, the more fit the agent, the more likely it is to reproduce and so pass on its evaluation network settings. The question the author asks is: will evaluation networks evolve to produce better fitness predictions through inheritance? The answer is yes, but with a surprising twist. The author's simulations make the perplexing discovery that, although the average fitness of the agents rises significantly in 200 generations, the success at predicting the next state by the model network decreases and the evaluation network acquires an inaccurate mapping from states to fitness. A careful analysis reveals that agents with slightly distorted representations are actually more fit because they allow encoding of more flexible, adaptive strategies. Thus, these simulations suggest that cognitive systems that represent false beliefs may well be adaptive.