As part of the Canadian Distributed Mentorship Program, here is where I will be posting my journal entries relating to the work done.

This week, the Reasoning and Learning lab played host to the first INRIA-McGill Workshop. What this meant is that none of us worked on their actual projects in the second half of the week; instead, we attended several lectures, mostly focused around bisimulation and LMP's, while eating enough shortbread cookies to feed a small subsaharan village. Oh and coffee. Lots and lots of coffee.

Early in the week, Eric and I started actual work on the Nortel project. Pablo had previously run some scripts to isolate the Code Blue events from the monstruously large simulation data. However, after we played around with the Weka trees on the data, we realized that the 30 minute window we had allocated for the Code Blue wasn't enough to witness its completion. We did, however, find patterns in the people that were assigned to a Code Blue emergency, which promised that the decision trees would help with finding the rules. Using some fancy awk scripting that I like to steer widely clear of, Eric generated some files that described where everyone in the hospital was, what their status was, and whether it was interruptible. Running a C4.5 tree, however, was absolutely dissapointing. The tree assigned one anaesthesist, and then went on its merry way. Tree, no banana bucks for you!

After a short e-mail exchange Eric and I thought that we ought to find a way to bound this data. The reason while the tree failed was because the number of people not assigned to the emergency(the negative examples) was significantly greater than the ones that did(the positive examples), so the tree algorithm cannot build a good rule. Armed with more awk scripting, as well as Dijkstra's algorithm (most commonly known for being the Spelling Bee nightmare, and also, in its days off, for efficiently solving shortest-path graph problems), we added distance-to-the-emergency information to the tree. This improved things, but very slightly, and the tree failed again.

And that's when we stopped growing productivity in the lab, and instead turned to the shortbread cookies.
Wikipedia can tell you about bisimulation much better than I can. It blew my mind. The really dumb down summary would be that a bisimulation is an equivalence relation between two state-transition systems (like a Markov Decision Process, or a Labelled Markov Process), that tells you how similar one system is to the other, and whether they behave identically. That's when Analysis IV, Sigma Algebras and the odd Analytic Space take the scene, and I am left to my own devices, namely shortbread cookies.