Tony Cassandra gave a nice talk Friday for FAI. I had had an informal intro to POMDPs earlier (from Tony), but it was nice to see it presented more formally. The issue of how to reason and act with hidden state (i.e. in “partially observable” domains) is going to be very important if we want intelligent agents that deal autonomously with people and other agents, since by default the other agents’ internal stage (what we’re thinking) is hidden.
One important thing is that all the work that Tony presented, and (I think) all the current work on POMDPs assumes that you have a (probabilistic) model of the environment. How an agent can learn such a model is still totally an open question. Agents that can build and/or adapt their world model on-line are a great field of research (if I do say so myself. ;-) )
Apart from the great content, Tony’s talk was interesting on a couple of other levels. First, instead of using powerpoint, or any kind of slides or “slideware”, he just did the visuals for his talk as one big web page. This is interesting because the day before the talk I received my copy of Edward Tufte‘s essay The Cognitive Style of Powerpoint, which I’ve mentioned before. Tufte describes in great detail the ways in which PowerPoint is bad for communicating complex ideas. Particularly bullet lists of short sentence fragments and breaking up the presentation into slides. Tony’s talk web page has breaks in natural places for the subject matter, and he just scrolled down as needed while talking. He also has (mostly) full sentences describing ideas and few bullet lists.
The other reason it was interesting was that to present the talk he just put up the web page on his server. Then when he arrived he used my laptop connected to the CS dept wireless net. Not only did everything work flawlessly, but I never had any doubt that it would. How’s that for ubiquitous computing?