Friday, July 15, 2011

#ALGORITHMS: "Smarter AIs Read the Manual"

MIT's AI read the manual to learn how to play "Civilization" game.
Artificial intelligence has made great strides in solving particular problems, such as IBM's Watson, which learned to beat humans at the game Jeopardy. Unfortunately, dozens of human experts needed to read and digest rules of the game in order to code them into algorithms. Now a group at MIT wants to change the hand coding of algorithms by allowing an AI to read and digest the rules all by itself.

Massachusetts Institute of Technology researchers are crafting the "last AI" for playing games against humans--a set of algorithms that can read the manual for a game and then produce code that allows it to square off against rivals and even learn as it goes along. If successful, the team wants to install its machine learning algorithms into robots that learn from the ultimate rulebook--directly from interactions with its environment.

Artificial intelligence has not even come close to the original vision of providing humans with tireless software assistants capable of helping us with ad hoc tasks. The dream has been realized for specific cases, such as IBM's Watson. Watson was handcrafted to play the TV game show Jeopardy, and is currently being used for mining the exabytes of data in medical databases. Similar AI's have also been handcrafted to beat humans at chess, poker and many other computer games.

Researchers are applying their new AI learning approach to Sid Meier's Civilization V.

Unfortunately, to perform in this manner, extensive hand-coding efforts must be undertaken, sometimes by legions of programmers. In these situations, the programmers must craft a set of algorithms that understand the rules of the game and have some way to develop strategies to win. Such efforts have been limited in the past. For example, chess games typically have "play the computer" modes, but these merely use brute-force look-ahead tables to make moves that are the most likely to succeed. Most other computer-modes in games use hand-coded strategies that human experts have encapsulated into rule-sets that are then applied when those circumstances are recognized.

Now researchers in MIT’s Computer Science and Artificial Intelligence Lab believe they have a better way: create an AI that can generate its own algorithms after reading the manual for a specific task. The group, led by professor Regina Barzilay, got started by crafting an AI that could read instructions for installing new software and then generate scripts that accomplished the task autonomously. Assisted by David Silver, Royal Society University Research Fellow at University College London, and MIT doctoral candidate S.R.K. Branavan, the group is now tackling board games.

The first game the group has taken on is Sid Meier's Civilization V in which players try to develop a city into an empire. Using their AI to read the Civilization manual and augment its more traditional machine learning algorithms, the MIT group's new system was able to increase its victory rate over humans to 79 percent (over a previous high of 46 percent without its manual-reading ability).

Next, the team is aiming to generalize their results for robots in unstructured environments. There, the robot will learn the meanings of words the same way that children do--by touching the hot stove and burning their plastic-covered metallic fingers--that is, through exploratory interactions with their environment. Funding was provided by the National Science Foundation.
Further Reading