May 29, 2010
Yesterday’s Shopping Bot of Tomorrow

There was a time a dozen years ago where there was a pretty clear path forward for what the Internet would enable (throw an “e-” on that bitch!). You’d want to buy a camera. But different features are worth different amounts to you. Zoom? Format? Maker? And you only want one camera. Or maybe two? Complicated! The obvious thing to do is just to incorporate your preferences into a bidding bot, who would then navigate the information superhighway on your behalf. Of course, your bidding bot would interact with the store’s selling bot, but also everyone else’s bidding bot. What would be the result of this interaction? What’s the correct way to structure this?

There is now an awfully extensive literature behind all of this. It was, I believe, the original “real world link” behind the presence of game theory in the multi-agent systems field, and the cause of the quaintly named ACM Conference on Electronic Commerce (EC). It’s also the reason my research group has the awful name Agent-Mediated Electronic Marketplaces Lab (hey, if we were all using shopping bots it would sound a lot better).

It’s ironic though that in the real world of Electronic Commerce, what’s succeeded is precisely the opposite. Instead of giving you a huge, expressive range of controls over items and prices, the cool, hip new services take a single thing for a single day and sells it for a single, really cheap price. Though Groupon is the hottest example, the real starter of the movement was Woot, which has been around since 2004. I remember when I first heard of it, the idea just seemed nuts. What kind of store sells only one thing, and changes that thing every day? Who would shop at a store that sold only a single external hard drive one day, and then only a single bluetooth headset the next day? Enough people to make a whole bunch of money, apparently.

So why have Wooty services succeeded and shopping bot services failed? I think it’s all symptomatic of one of our less visible challenges — that preference elicitation is incredibly hard. It’s of course hard computationally, with exponential state spaces, but the most challenging issue is that standard utility theory doesn’t do a great job of actually modeling the way people really behave. (You can either believe that the theory is defective, or that people are defective.) So even if you do manage to do it “right”, people will still get angry at you. NRMP Lawsuit, anyone? So what’s emerged from the Internet morass is the least eliciting mechanism possible: your preferences are just a single bit and the mechanism posts a price. Would a shopping bot model be more efficient? Yes, but only if your model is right — and we’re at the point now in “Electronic Commerce” where the failure of shopping bots to emerge argues that the model is wrong.

4:59am  |   URL: http://tmblr.co/ZtlAMycLpWe
(View comments
Filed under: AI 
Blog comments powered by Disqus