It was headline news in May of 1997 when the IBM computer ‘Big Blue’ beat chess champion Gary Kasparov. So why is it that Kumar Chellapilla of the University of California in San Diego and David Fogel of Natural Selection Inc. have written two papers1,2 in which they suggest that this seeming triumph “is testament to the limitation” of the very idea of artificial intelligence?
Artificial intelligence is often hailed as a new development, but the year 2000 will see its fiftieth birthday. In 1950 Bell Laboratories engineer Claude Shannon wrote the first chess-playing program for a computer. This allowed computers to hatch out of the shell of numerical calculations and perform tasks that could be said to require logical ‘thought’. Shannon showed3 that the game of chess could be reduced to numerical manipulation through the invention of a ‘minimax’ algorithm that allowed a computer to choose a good move without having to evaluate all the states in the intervening moves. This became proof that the number-crunching abilities of computers could be used to perform tasks which, if done by humans, would be said to require intelligence.
Chellapilla and Fogel's critique is based on an approach that avoids the need for a programmer to work out the game-playing program. It uses evolution in artificial neural networks which, they argue, is an advance both for artificial intelligence and the design of such networks. The proposed artificial evolution encourages the survival of those systems in a batch that exhibit success in beating artificial opponents. New batches are derived from successful networks and the process is repeated until some good playing strategies have evolved. Such strategies have been shown to beat human experts. The game in Chellapilla and Fogel's case is not chess but checkers, or draughts as it is called in Britain. The authors point forcefully to the fact that the evolutionary process actually creates its own intelligence rather than relying on that of a programmer. Referring to the method that eventually defeated Kasparov, they write: “Every ‘item’ of knowledge was pre-programmed. In some cases the programs were even tuned to defeat particular opponents…”.
Although neural networks were first suggested in the early 1940s, it is only in the past 15 years or so that they have been used with some success as systems that can learn to recognize patterns in a variety of scientific and industrial applications. Conventionally they consist of layers of cells (artificial neurons) where interconnections between neurons in adjoining layers are made through ‘weights’, while the body of the cell performs summations that cause it to ‘fire’ at an output (the axon) when the incoming weighted activity overcomes some mathematically specified gradient function. The weights are thought to resemble the action of synapses in real neuronal cells found in biological brains, where neurons connect and learning takes place through the alteration of such weights.
So an artificial network (which can have one of many popular topologies) learns to approximate an input–output function by being given examples of the in- and out-values of such a function. It does this through one of a variety of weight-adjustment procedures known as learning algorithms. The learned tasks can be as esoteric as recognizing faces where it would be hard for a programmer to work out a priori a function that would distinguish between such patterns. However the techniques used by Chellapilla and Fogel do not make use of learning algorithms such as these, but arrive at suitable synaptic-weight functions through an evolutionary process.
They use a population of 15 neural networks with randomly selected weights and the ‘offspring’ of such parent networks containing a variation of the parents' weights. The task for the network is to evaluate board images, while the playing proceeds along Shannon's minimax lines. All parents and offspring play five games against opponents randomly selected from their midst and are awarded points based on their successes and failures. The 15 most successful networks are then chosen to be the new generation, offspring are produced and the process repeats. One hundred generations take about two days of computing on a 400 MHz Pentium machine. After 250 generations the best network is chosen to play human all-comers on the Internet, who in turn are evaluated in terms of the excellence of their playing on a national categorization. The network was able to achieve a top categorization (that is, ‘A’) on this scale, to defeat two ‘expert’ category players and to draw with a ‘master’.
Chellapilla and Fogel stress that evolutionary methods can develop machine intelligence without being programmed, and demonstrate1 some of the generality of this principle in other games such as tic-tac-toe (noughts and crosses) and those involving a mix of collaborative strategies between the players (the so-called Prisoner's Dilemma). All of this shows that organisms in nature achieve what is called intelligence through a fascinating mix of evolution, adaptation and learning — with the possibility of inspiring those interested in computation not only to build smarter machines but also to get a better understanding of the nature of intelligence itself.
Chellapilla, K. & Fogel, D. B. Proc. IEEE 8, 1471–1496 (1999).
Chellapilla, K. & Fogel, D. B. IEEE Trans. Neural Networks 10, 1382–1391 (1999).
Shannon, C. E. Phil. Mag. 41, 256–275 (1950).
About this article
Expert Systems (2007)
Proceedings of the IEEE (2004)