FUNDAMENTALS OF GAME DESIGN, SECOND EDITION

Creating Artificial Intelligence

In 1959, IBM scientist Arthur Samuels devised a program that played checkers (Samuels, 1959). The program could also learn from its mistakes, and eventually it became good enough to beat expert human players. Much of the earliest research on artificial intelligence and games was of this sort as computer scientists tried to create artificial opponents that could play traditional games as competently as humans could. Artificial intelligence (AI) lets us play multiplayer games even when we don't have other people to play with.

However, AI brings considerably more to video gaming than artificial opponents for traditional games. Game developers use AI techniques for the following:

■ Strategy. This means determining the optimal action to take by considering the possible consequences of a variety of available actions. Samuels's checker-playing program did this, but checkers is a game of perfect information, which means there is no hidden information and no element of chance. Modern video games usually have both hidden information and a large element of chance, so a strategy is more difficult to compute.

■ Pathfinding. This means finding the most advantageous routes through a simu­lated landscape filled with obstacles.

■ Natural language parsing. Despite decades of research, computers still cannot understand ordinary written or spoken language well, but researchers are still very interested in using it for games. When this problem is solved, players will be able to give commands using natural sentences.

■ Natural language generation. Video games currently produce language by playing combinations of previously recorded phrases or sentences. At the moment, they cannot generate language on their own. In time perhaps they will, which will make simulated people seem far more realistic. In the meantime, games use AI to select a sentence from their library of prerecorded material that is most appropriate for the current game situation.

■ Pattern recognition. This valuable technique has numerous applications including voice recognition, face recognition, pattern detection in ongoing pro­cesses, and pattern detection in player behavior. Human poker players use pattern recognition to establish a correlation between their opponent's behavior and their
opponent's cards, which players can use to their advantage later. Eventually, a com­puter might be programmed to do the same thing.

■ Simulated people and creatures. Many games use simple AI techniques to cre­ate a behavioral model for simulated people or creatures. The simulated character seems to respond intelligently to the human player's actions, at least within certain limits. The models are seldom complex, and a player can usually tell the difference between a simulated person and a real one within a few minutes. Simulating human beings is the most difficult and also the most important problem in game AI research.

Most current video games do not, in fact, contain much real AI. The point of video games is to entertain, not to simulate intelligence in depth, so they usually contain just enough AI to make the player feel as if the software is reasonably smart. The players—who are already immersed in the make-believe world anyway—are often happy to give the game credit for intelligence that it doesn't really possess. However, we can't afford to rely on this. Sometimes the AI fails, and the game breaks the players' immersion by doing something startlingly unintelligent; we need to continue to improve our AI so this doesn't happen. The future looks promising. Technological advances in both hardware and software are allowing designers to create increasingly sophisticated artificial opponents and simulated characters. Artificial intelligence is one of the most important areas of research in game development.