Computer reads manual, wins Civ

A group of researchers have developed a computer system that learns how to play Civilization, using only the game's manual and onscreen feedback to develop a strategy that saw it win 79 per cent of its games - a development which could greatly aid the videogame AI design process.

The researchers, from MIT's Computer Science and Artificial Intelligence Lab, give their computer system no prior knowledge of the game beyond a list of inputs - left and right clicks, and movement of the cursor - or the English language. It has no knowledge of which actions corresponded to words in the instruction manual, nor what objects in the game world represent. The only feedback it has is on-screen text, and whether an action results in a win or a loss.

As such, its behaviour at the outset is entirely random. As it experiments, it takes text from on-screen feedback and searches for the same words in the instruction manual. It searches the surrounding text for associated words and develops hypotheses: those that are met with success are retained, and others are discarded. Over time, it developed a play strategy that saw it win almost eight out of every ten matches in Sid Meier's classic strategy game.

The implications for game development are significant. Programmers working on complex games like Civilization have to create singleplayer AI algorithms themselves, a laborious, time-intensive process. The researchers believe that their system would not only automate that process, and speed it up, but yield better results.

"Games are used as a test-bed for artificial intelligence techniques simply because of their complexity," said S R K Branavan, one of the researchers. "Every action that you take in the game doesn't have a predetermined outcome, because the game or the opponent can randomly react to what you do. So you need a technique that can handle very complex scenarios that react in potentially random ways."

Regina Barzilay, associate professor of computer science and electrical engineering, explained that game manuals have "very open text. They don't tell you how to win. They just give you very general advice and suggestions, and you have to figure out a lot of other things on your own." For Barzilay and her researchers, games are "another step closer to the real world."

It's a wonderful example of games aiding scientific research, and the fruits of that research potentially aiding the development of the very same type of game. Indeed, the implications spread far wider than our preferred hobby: Barzilay and her students are now in the process of adapting their language-learning, meaning-inferring algorithm for use in robotics.

Source: MIT