Friday, September 14, 2007

2003 Humans and computers compete in virtual creature game

Humans and computers compete in virtual creature game

  • 17:00 05 December 2003
  • NewScientist.com news service
  • Will

An online game that lets contestants build and race virtual beasts is being used to pit humans against a variety of artificial intelligence algorithms.

The objective of Sodarace, which started at the end of November, is to construct a two dimensional creature that can travel over a certain type of terrain in the shortest possible time.

Each creature is constructed of "mass", muscles", "limbs" and "joints" which control the way it moves. These creations can then be raced over a piece of terrain. Creatures can have many limbs or none at all and can walk, wriggle or even jump along.

It is relatively simple to construct a creature by hand. But the game has been written so that a creature's key parameters can easily be fed into another computer program and artificial intelligence (AI) programmers are being invited to take part.

So far, Sodarace has attracted thousands of contestants from around the world. These include hobbyists and professional AI researchers. In the first round, a human player was able to outwit competing computer algorithms to develop the fastest creature.

'Offspring' algorithms

Peter McOwen, an AI expert at the University of London who helped organise the project, says the game provides useful way of comparing different AI approaches. "If you have an AI algorithm, you want to know how good it is at a known problem" he told New Scientist. "This is a fun way to show that your algorithm is better than someone else's."

McOwen's group is using genetic algorithms to "evolve" the fastest artificial creature. This involves creating a population of algorithms that compete with one another to make the best animal. The algorithms that work best are combined to create "offspring" algorithms that also compete.

Other research groups are using entirely different AI approaches. For example some involve neural networks - systems which simulate neurons in the human brain to tackle a problem. Other AI programs use or a process known as "simulated annealing" which whittles down strategies to find the optimal solution using the least effort.

The project was organised by University of London with a UK-based multimedia design company called Soda.

No comments: