Feature: 'Evolving Pathfinding Algorithms Using Genetic AI'

In today's main Gamasutra technical feature, Cubism author Rick Strom explores the possibility of improving in-game AI by exploring an advanced computer very close to home; the human brain, using what's called genetic programming. From the feature's introduction: "Pathfinding AIs tend to fall short in one of two ways: either they are so incredibly effective and efficient that they don't convey intelligence as much as omniscience, or they are so incredibly inept that human observers can't help but feel frustration watching them. The first case is exemplified by an agent which ignores locally acceptable paths in favor of the ideal global path, thereby giving away the secret that it has knowledge of the entire map. The second case can be seen in a wide variety of stupid behavior, but my favorite example is the agent which travels along a river to avoid the relatively small penalty of crossing water, only to plow through an enemy camp and ultimately meet its doom. Anyone familiar with real-time-strategy gaming has no doubt met this breed of pathfinder before. The problem, assuming we choose not to endow our pathfinders with God-like knowledge of the Universe, is that we are trying to simulate the sort of decision making employed by a computer that we still don't understand - the human brain. Rather than trying to hard code something this complex, perhaps we should be allowing our algorithms to develop over thousands and thousands (or millions and millions) of generations, using the same evolutionary mechanism that nature has relied on all along." You can now read the full Gamasutra feature on the subject, including plenty more on the concepts and reality of genetic programming in relation to AI (no registration required, please feel free to link to this feature from external websites).