Skip to main content

Neural Networks within Games

An actual neural network is made up of neurons that are connected in the nervous system of every human being.  However a neural network can be simulated by nodes and links in-between the nodes.  This type of network almost electrically controls human behavior within the brain, so a simulation could in effect control the behavior of artificial intelligence.  Artificial intelligence within video/computer games is an interesting place to find implementations of neural networks.  Unfortunately many do not use implementations of neural networks, however this might change in the future.  Most artificial intelligence within games implement decision trees or finite state machines to process their logic and produce outcomes for their behavior.

Usually the artificial intelligence within a game does not react to the user.  Yes, the artificial intelligence does interact with the user, otherwise most games would be exceedingly boring.  However, not many actually react to moves that the player makes.  They are all pre programmed to react to specific moves or signals that the player chooses to play throughout the game.  However, if the player comes up with a new move that wasn’t thought of by the original programmers (especially within indie video games that do not have the budget of larger corporations) it would be very easy to kill off all of the artificial intelligence and beat the game.  Neural networks can be used to train the artificial intelligence to learn new moves or ideas within the game.  Game design companies have been trying to implement these since the early 2000s, and have been gaining greater success.  The reason why these were so hard to implement at the beginning is that they are very memory intensive and it would be hard to train all of the AI at runtime without the game being a bust.  If all of the AI had to be trained by the user at runtime using a generic genetic algorithm the game would be boring and would not change much throughout the levels due to very very slow changes.

There is a lot of randomization within a neural network which does lead to an equilibrium, but it might not always be unique.  Equilibrium within an AI gaming network would occur when all of the AI were sufficiently trained to react to some arbitrary movement by the user.  Neural networks have had greater success within the past few years within the gaming industry, however finite state machines and decision trees are still preferred.  It will be interesting to see what happens with neural networks in the future.


Leave a Reply

You must be logged in to post a comment.