We had two final project presentations today for CIS 730 (Artificial Intelligence). Both were among the five who worked on the Roguelike game Angband. Our goal was to look at specific behaviors and improve them.

The first student, Dave Lupo, wanted to improve the tendency of the BenBorg (by Ben Harrison) to be a shopaholic. He trained a feedforward artificial neural net (ANN) using backpropagation to compute a better "dive motivator". This lowered the ratio of time in town vs. dungeon, and he found that increasing the ratio of "time in the dungeon" to "time in town" increased survivability.

Dave plotted the "time in town vs. time in dungeon" curve for 13 characters before his improved dive function, and 14 characters after, and found that they did have higher XP-to-move ratios. He speculated that they had higher survivability as a result, though these results were inconclusive. (I suggested that he look at the slope of the line to see if ' it really improved survivability.)

Now, here's the funny part. The points were all at time of character death, because he lost most of the characters at low levels, but I was sure he didn't lose them all by level 14, so I asked him what the rightmost point was. "Oh, that's time of death after 150000 turns". I asked, "what do you mean, after 150K turns?" He replied that to impose a time limit, he didn't just end the borg run at 150K; he sets "target level = 99" so that it essentially goes: "Morgoth... I'm comin' to get you!" and commences a Rambo-esque death dive!

--
Banazir

An introductory course in machine learning for development of intelligent knowledge based systems. The first half of the course will focus on basic taxonomies and theories of learning, algorithms for concept learning, statistical learning, knowledge representation, pattern recognition, and reasoning under uncertainty. The second half of the course will survey some basic topics in combining multiple models, learning from time series, learning to reason, and selected applications in knowledge discovery and data mining, especially in bioinformatics.

The course will include several written and programming assignments and a term project option for graduate students. Ancillary readings will be assigned; students will write a brief synopsis and review for one of these papers every other lecture.

This will be my sixth offering of machine learning (the first five being in 1999, 2001, 2002, 2003, and 2005). This time I'm cross-listing it with my Advanced AI course (CIS 830), and am giving approximately equal time to graphical models, genetic and evolutionary computation (especially genetic programming, but with some genetic algorithms coverage), and artificial neural networks.

--
Banazir

UK Trip 2005: A Tronkie Travellogue
Day 12: Edinburgh, Scotland to London, England (IJCAI Conference)

( Farewell to Caledonia )

--
Banazir

UK Trip 2005: A Tronkie Travellogue
Day 11: Edinburgh, Scotland (IJCAI Conference)

( Live from Canadia, it's Thursday Night! )

--
Banazir