www.sciencemag.org SCIENCE VOL 325 17 JULY 2009 277 PERSPECTIVES S ince its creation in the early 1990s, the World Wide Web has evolved from its initial, static Web sites to the dynamic, interactive Web sites of today. But whereas existing Web sites merely respond directly to user input, there is growing interest in making them adaptive through the use of computa- tional intelligence. A promising approach for this involves a family of optimization tech- niques called evolutionary algorithms. A typical evolutionary algorithm ( 1) starts with an initial population of randomly generated, digital solutions to a problem (“individuals”). These individuals are eval- uated with a user-supplied fitness function; on the basis of their fitness scores, better individuals are stochastically selected to act as “parents.” Either one parent is cop- ied while making a small change to it (muta- tion), or parts of two parents are combined to make a new individual (recombination). This breeding process is repeated for a fixed number of evaluations or until the problem has been solved. Since the 1990s, evolutionary algorithms have been applied to architectural problems from arch dams and suspension bridges to building plans. In industrial and engineering design, they have found use in color design for knitwear, shape design for scissors, and car body styling, as well as for creating com- plex devices such as gyroscopes and wind turbines. Examples are the nose cone of Hitachi’s Series N700 Bullet Train ( 2) and the communications antennas for the space- craft in NASA’s ST-5 mission, a test mission to validate new space technologies and study the magnetosphere ( 3). Traditional evolutionary algorithms optimize against explicit fitness functions, but problems involving taste or aesthetics cannot be easily reduced to a mathemati- cal function of goodness. Instead, a human user can perform evaluation manually, as first proposed by Dawkins ( 4). Running on a standard PC, the first interactive evolution- ary algorithm showed the user computer- generated images, of which the user would select one as the parent for the next genera- tion. By iteratively selecting images on the basis of aesthetics, the algorithm produces more and more visually appealing images over time. This has become the standard interface for interactive evolutionary algo- rithms (see the figure). Since then, interactive evolutionary algo- rithms have been used in various human- computer interactive design systems. Most applications are visual, such as the evolution of images ( 5), three-dimensional shapes ( 6), and architectural forms ( 7), but they have also been used for musical tasks, including sound synthesis and composition. For exam- ple, GenJam is an interactive evolutionary algorithm for real-time jazz improvisation ( 8). Along with its creator, Al Biles, it forms a virtual jazz quintet that has performed at more than 100 private receptions. The first Web browser that supported images (Mosaic) was released in April 1993. Soon afterward, the first online interactive evolutionary algorithms appeared. The Inter- national Interactive Genetic Art 1 (IIGA1) ( 9) and its successor, IIGA2, had more than 100,000 visitors who collectively created thousands of images over a period of 4 years. In an early commercial application, Affinnova (www.affinnova.com) has used an interac- tive evolutionary algorithm–based system to design product packaging since 2000. Nym- bler (www.nymbler.com) allows users to evolve baby names instead of images. A key challenge for interactive evolution- ary algorithms is user fatigue ( 10). For typi- cal noninteractive evolutionary algorithms, tens of thousands of evaluations are needed to achieve interesting results—orders of magni- tude more than can be expected from a single user. On the Web, many users are available, but even this multiplier effect may not over- come user fatigue: Because the interactions are distributed in time, no single user is likely to experience evolution at a sufficiently fast pace for it to be interesting. Toward a Smarter Web COMPUTER SCIENCE Gregory S. Hornby 1 and Tolga Kurtoglu 2 Interactive evolutionary algorithms are increasingly implemented in Web sites to respond to user preferences. 1 University of California at Santa Cruz, University Affili- ated Research Center, Mail Stop 269-3, Moffett Field, CA 94035, USA. 2 Mission Critical Technologies Inc., NASA Ames Research Center, Moffett Field, CA 94035, USA. E-mail: gregory.s.hornby@nasa.gov; tolga.kurtoglu@nasa.gov A B Learning by doing it. An explicit interactive evo- lutionary algorithm presents the user with several computer-evolved design options (A). On the basis of the user’s selection (red outline), a new set of evolved designs is presented (B). Published by AAAS on July 28, 2009 www.sciencemag.org Downloaded from