The phenomenon of learning generalization - where an organism repeats

behavior learned in response to one stimulus when presented with a perceptually

similar stimulus - has been well documented in a variety of animals.

I argue that evolutionary game theory can help explain the prevalence of

this type of learning behavior by showing how and when generalization can

outperform other strategies in situations where there are payoff similarities

between states.

Jäger (2007) introduced Sim-Max games, a variation of the standard

Lewis signaling game where the state space is endowed with a metric that

captures a similarity relation over states of the world. This added structure

is manifested in payoffs that reward behavior in both the ideal state for that

behavior as well as similar states. A modication of this game can be used

as a good model to explore the success of learning generalization in single

organism situations.

I show that in these games learning generalization can sometimes outperform

simple reinforcement learning. However, it does not do so in all

cases. My results highlight an interesting tension. The strategies developed

by generalizing learners are necessarily imprecise, and thus perform less well

than ideal strategies in these games. However, learning generalization allows

actors to develop a fairly successful strategy very quickly. I show that

generalization can be expected to evolve in cases where organisms need to

learn how to act in many different states over a short time scale.

References

Jager, Gerhard (2007). "The evolution of convex categories." Linguistics and

Philosophy, 30, 551-564.