Calling all agents
First published in Personal Computer World, June 1996.
IT IS 10 YEARS SINCE MARVIN MINSKY, founder of MIT's Artificial Intelligence Laboratory, proposed in his ground-breaking book The Society of Mind that intelligence can emerge from non-intelligence.
In Minsky's view, the human mind is made of up of many processes he calls 'agents', each of which performs a single very simple task requiring no innate intelligence whatsoever. But connect enough of these agents together in 'certain very special ways', as Minsky says, then true intelligence somehow emerges. These agents of the mind are the inspiration for a new kind of software design, which may hold the key to the future of computing.
Just like the tools we know and love today -- our spreadsheets, editors, word-processors and web browsers -- software agents will perform tasks for us, but will do so much more powerfully. They will be goal-directed, may operate autonomously, learn from their previous experiences, be reactive to their execution environments, and collaborate with other agents. The claim is that recognisably 'intelligent' behaviour will result.
Suppose you decide to buy a new colour printer. Instead of ploughing through pages of ads, you summon a suitable software agent and tell it: "Find the fastest colour printer under L400, with the best warranty and upgrade deal, and the best reviews". Off your agent goes, marshalling the appropriate internet search agents to scan on-line dealers' catalogues, discussions in appropriate newsgroups, and relevant web sites, comparing features and opinions, and eventually reporting back to you a summary of its findings. You then scan the list, click 'yes' against your choice, and off goes another agent, ordering the item with your credit card, and finally telling your 'personal finance' agent that you've made a transaction.
It sounds far-fetched, but prototypes of such agents already exist. Andersen Consulting's 'BargainFinder', for example, allows comparison shopping for CDs. You supply the title of the CD, and BargainFinder visits nine retailers on the web on your behalf, returning a summary of prices and delivery charges.
The current buzz is to call such agents 'intelligent'. But are they? 'Intelligence' is a loaded term: one person's intelligent agent is another's humdrum software tool. Nevertheless, agents are now a hot topic of research in the Computer Science and Artificial Intelligence communities, and intelligent or not, it is certain that agents of one kind or another are set to play a prominent role in information retrieval and filtering.
Traditionally, information databases have been the domain of computing professionals, but in the last couple of years, anyone with a modem now has access to the largest world-wide distributed database of all time: the web. The problem of making sense of the unimaginable amount of information out there is a very hard one. Current strategies for indexing the web are rather naive, and rely on software robots crawling around the world, accumulating and indexing information as they go. The resulting databases are *huge*. Lycos, for example, currently claims to store 32 gigabytes of indexed web data. Such databases are monolothic, and will not scale well as the size of the web increases. In the future, we should expect to see a proliferation of agents compiling smaller, specialised and interconnected databases.
Currently, when seekers of information perform keyword-based searches on web databases, these are typically one-off operations. An agent, however, may use its experience of previous searches, and -- crucially -- work in collaboration with other agents. The University of Massachusetts 'searchbots' are a step in this direction, offering goal-directed searching with multiple agents performing parallel searches using standard web tools, filtering the results appropriately.
A similar agent, currently in beta-test, is IBM's 'infoSage', to which you submit an 'interest profile', which is used to filter news and information sources on the web, emailing you twice a day a 'personalized newsletter' of snippets relevant to your interests. As with many such systems, it would currently be inappropriate to call infoSage 'intelligent', since its filtering algorithms are somewhat inscrutible: while it often works well, some users report receiving articles which appear to have nothing whatsoever to do with their interest profiles. And one can't help wondering just what IBM is doing with all those profiles they collect...
Although the ideas in agent research are exciting, some people find agents decidedly unattractive. They worry that ultimately agents may do too much, leaving their users confused about what is being undertaken on their behalf, leading to alienation. There is also a view that the claims for intelligent agents will lead to unfulfilled expectations, rather like the failure of Artificial Intelligence research to live up to its promises in the 1970s. Surprisingly, one of the fiercest critics of agents is Jaron Lanier, a pioneer of virtual reality.
Intelligent agents are not yet here, but the chances are good that they will arrive soon. And -- so long as we can trust them -- we shall one day wonder how we ever managed without them.
Toby Howard is a lecturer at the University of Manchester.