Chapter 14. Neural Networks

Our brains are composed of billions of neurons, each one connected to thousands of other neurons to form a complex network of extraordinary processing power. Artificial neural networks, hereinafter referred to simply as neural networks or networks, attempt to mimic our brain's processing capability, albeit on a far smaller scale.

Information, so to speak, is transmitted from one neuron to another via the axon and dendrites. The axon carries the voltage potential, or action potential, from an activated neuron to other connected neurons. The action potential is picked up from receptors in the dendrites. The synaptic gap is where chemical reactions take place, either to excite or inhibit the action potential input to the given neuron. Figure 14-1 illustrates a neuron.

Figure 14-1. Neuron

The adult human brain contains about 10₁₁ neurons and each neuron receives synaptic input from about 10₄ other neurons. If the combined effect of all these inputs is of sufficient strength, the neuron will fire, transmitting its action potential to other neurons.

The artificial networks we use in games are quite simple by comparison. For many applications artificial neural networks are composed of only a handful, a dozen or so, neurons. This is far simpler than our brains. Some specific applications use networks composed of perhaps thousands of neurons, yet even these are simple in comparison to our brains. At this time we can't hope to approach the processing power of the human brain using our artificial networks; however, for specific problems our simple networks can be quite powerful.

This is the biological metaphor for neural networks. Sometimes it's helpful to think of neural networks in a less biological sense. Specifically, you can think of a neural network as a mathematical function approximator. Input to the network represents independent variables, while output represents the dependant variable(s). The network itself is a function giving one unique set of output for the given input. The function in this case is difficult to write in equation form and fortunately we don't need to do so. Further, the function is highly nonlinear. We'll come back to this way of thinking a little later.

For games, neural networks offer some key advantages over more traditional AI techniques. First, using a neural network enables game developers to simplify coding of complex state machines or rules-based systems by relegating key decision-making processes to one or more trained neural networks. Second, neural networks offer the potential for the game's AI to adapt as the game is played. This is a rather compelling possibility and is a very popular subject in the game AI community at the time of this writing.

In spite of these advantages, neural networks have not gained widespread use in video games. Game developers have used neural networks in some popular games; but by and large, their use in games is limited. This probably is due to several factors, of which we describe two key factors next.

First, neural networks are great at handling highly nonlinear problems; ones you cannot tackle easily using traditional methods. This sometimes makes understanding exactly what the network is doing and how it is arriving at its results difficult to follow, which can be disconcerting for the would-be tester. Second, it's difficult at times to predict what a neural network will generate as output, especially if the network is programmed to learn or adapt within a game. These two factors make testing and debugging a neural network relatively difficult compared to testing and debugging a finite state machine, for example.

Further, some early attempts at the use of neural networks in games have tried to tackle complete AI systems—that is, massive neural networks were assembled to handle the most general AI tasks that a given game creature or character could encounter. The neural network acted as the entire AI system—the whole brain, so to speak. We don't advocate this approach, as it compounds the problems associated with predictability, testing, and debugging. Instead, just like our own brains have many areas that specialize in specific tasks, we suggest that you use neural networks to handle specific game AI tasks as part of an integrated AI system that uses traditional AI techniques as well. In this way, the majority of the AI system will be relatively predictable, and the hard AI tasks or the ones which you want to take advantage of learning and adapting will use specific neural networks that were trained strictly for that one task.

The AI community uses many different kinds of neural networks to solve all sorts of problems, from financial to engineering problems and many in between. Neural networks often are combined with other techniques such as fuzzy systems, genetic algorithms, and probabilistic methods, to name a few. This subject is far too vast to treat in a single chapter, so we're going to narrow our focus on a particularly useful class of neural networks. We're going to concentrate our attention on a type of neural network called a multilayer, feed-forward network. This type of network is quite versatile and is capable of handling a wide variety of problems. Before getting into the details of such a network, let's first explore in general terms how you can apply neural networks in games.