Summary

This chapter used two perceptrons to increase the realism of the aiming in animats, but also to allow them to deal with these errors effectively—as human players do:

• To improve the realism of the aiming, a perceptron learns to turn smoothly by approximating a more extensive model of aiming errors. This is learned offline using a batch training algorithm.

• To increase the effectiveness of the aiming, but still behaving realistically, another perceptron approximates a solution that compensates for the aiming errors. This is learned online by gathering training examples.

Because of the simple nature of the problems (easily approximable and tolerating suboptimality), the limitations of the perceptrons are not exposed. In many cases, it's preferable to spend additional time in AI design to be able to apply single-layer perceptrons. In some cases, however, this is not possible; the additional power of multilayer perceptrons is necessary—accompanied by the extra complexity. The next chapter covers the theory behind multilayer perceptrons, paving the way for a problem in Chapter 20, "Selecting the Target," that requires better capabilities to recognize patterns.