Imagining AI for a game usually conjures the image of simulated human players — robots designed to look and act like people that can stand in seamlessly for real human opponents. This expectation is natural and unsurprising; we humans have a strong innate tendency to anthropomorphize intelligent or semi-intelligent entities in any context, and the controlled environment of a game system provides an excellent framework for this projection. However, as an AI designer, you must sooner or later confront the realization that computers are not, and cannot be, effective simulated humans. To refer back to the quote from Dijkstra that opened this series, trying to squeeze a computer mind into the shape of a human mind is at best ineffective and at worst outright destructive to the success of your AI.
To illustrate this point, let us examine those ways in which the two kinds of minds encounter a game: a human player, and a computer player. Both should be equally capable of success and skillful play, but the avenues and mechanisms each will take to achieve that success are entirely different. Both will find greatest effectiveness under different circumstances, and will require different supports to make best use of their particular advantages. Acknowledging these areas where humans excel above computers, and vice versa, is the opening move in understanding how to architect an effective AI.
So what are some ways in which a human player will outperform a computer? Here are some suggestions:
Insight: Much has been made of the advantage the brain has as compared to a computer. It’s often described as being massively parallel: we can consider many different threads simultaneously, etc. What this means is that, when analyzing a problem, humans will draw upon an incredible array of inputs, cross-referencing them into ever-deepening layers of meaning to very quickly form a sophisticated understanding of the situation. Consider: without being conscious of it, a human player instantly and seamlessly incorporates all their past experiences with the situation, past experiences with similar situations that might be applicable, understanding of the nature of the system and its rules combined with observation of flaws or exploits in that system, understanding of their opponent or obstacles and their nature or behavior, deductive reasoning and predictions about how these forces will operate under potential future circumstances, and many more inputs, all in the same time frame. This leads to the “sudden flash” we think of as insight, and for which humans are renowned — especially in the context of a game, where the system operates with visible and predictable structure, this ability is tremendously powerful.
Emotion: At first, this may not seem like an advantage. We rightly associate rational thought with effective thought, and we know that emotion undermines rational thought. But like it or not, emotion is an important aspect of human decision-making, and one that games can and should account for. Emotion leads players to press on even against long odds, to try their luck after fortunate but unpredictable success, to stubbornly pursue goals and accomplishments they should have abandoned, and to lose their cool and expend energy or resources at wasteful rates (holding down the fire button until the gun goes ‘click’). While these irrational moves may not always be advantageous, they are an essential part of human play. What’s more, they are unpredictable and defy logic — something the cold, rational computer may have little or no capacity to respond to. And in any competition, any play that your opponent cannot handle suddenly becomes a lot more valuable.
Adaptation: Easily the most important point of power for a human player, and the very one that set our species apart all those millennia ago: humans are adaptable. If the circumstances change, we change with them. We learn. We improve. The human mind is constantly reevaluating, constantly probing and testing the system, searching for flaws, looking for new ways to understand it and new ways to manipulate it. And we do it fast. We are remarkably quick at noticing changes and assessing how to respond to them. We invent whole mechanisms of judgement, value, and prediction in a moment, and then re-invent them, again and again. Any player playing a game is a study in adaptation in progress. An AI can only ever express those strategies and tactics originally envisioned and implemented by its author. Cutting-edge academic AI scientists are trying to disprove this, but even with all they have managed to achieved, computers still can’t wholly reprogram themselves. A human player has no limit to how much she can grow, how far her understanding can reach. Humans are, and probably long will be, the most versatile, devious, and wily players possible.
CONTINUED in Part II: the tables turned, and a discussion of the ways in which computers outperform humans and how this should influence the architecture of game AI.