Chapter 19. Multilayer Perceptrons

Key Topics

• History of Perceptrons

• Model Overview

• Simulation

• Biological Parallels

• Training Algorithms

• Practical Issues

• Discussion

Multilayer perceptrons (MLPs) are another kind of artificial neural network, with layers of weighted connections between the inputs and outputs. The structure of MLP essentially resembles a set of cascaded perceptrons. Each processing unit has a relatively complex output function, which increases the capabilities of the network.

This chapter builds upon the information in Chapter 17, "Perceptrons," which covered single-layer perceptrons. This chapter covers the following topics:

• The history behind perceptrons, notably why multiple-layered models are necessary

• The representation of MLP, introducing the concept of topology and nonlinear activation

• The simulation of multiple layers of processing units, and how it differs from the single-layer variant

• The parallels between perceptrons with their biological counterparts: neural networks

• Methods for training MLP based on the concept of back-propagation of error

• Practical issues behind the training process, and the problems that occur with multiple layers

• The major advantages and disadvantages of perceptrons for game development

Perceptrons with multiple layers can recognize patterns in the gameplay, predict the outcome of a fight, and control the nonplayer character (NPC) movement. These tasks can be learned from a set of examples, and with potentially better performance than single-layer perceptrons.