MIT researchers have developed software that they claim has taught itself to play Civilization II - and to win most of the time, too.
In a recently published paper, the team explained how the software starts from nothing and learns to play by reading the manual, winning 80% of the time.
Beginning with a blank memory, and only able to move and click the mouse, the computer analyses the game screen. It picks out words such as "river" and "road", to search through the game manual.
Lead developer Regina Barzilay said they used the game's manual rather than a strategy guide because it gave the player "very general" advice, leaving a player to “figure out a lot of other things on [their] own”.
Finding the relevant search terms in the manual, the program then runs the paragraph through a sentence analysis algorithm, which it uses to form instructions.
With these instructions, the program can then form strategies. “Because of [a game's] complexity you need a technique that can handle very complex scenarios that react in potentially random ways,” said developer S.R.K. Branavan.
The researchers used the Monte-Carlo method, normally used for analysing financial simulations. It allows the developers' program to take a point in the game and test many possible responses. Keeping an eye on the game's score, it remembers which was the most beneficial decision and stores it in a database.
The amount of data stored by the system quickly becomes mind boggling. At every turn of its 200 games the computer would perform 500 possible actions, and “each roll-out is run for 20 simulated game steps before halting the simulation and evaluating the outcome,” the researchers said. Each of those 200 games are played for at least 100 turns, for a total of 200 million Civilization turns.
This article originally appeared at pcpro.co.uk
