ideas to try out

I hope these suggestions will inspire y’all to try new methods.

Jay : game learning : ideas

- AutoClass
AutoClass is a statistical classification system developed by the then-Bayesian Learning Group at NASA Ames. AutoClass C is a free implementation, written in ANSI C, with source code included. AutoClass is an unsupervised learner that discovers categories in data. In scientific data, it has discovered scientifically interesting categories. What might it discover in game data? Don’t overlook the value of unsupervised learning!

- the Baldwin effect
Applying evolutionary algorithms to neural networks and other learning systems.

- playout analysis
The generalization of backgammon rollout analysis, playout analysis is a search technique that concentrates on ultimate consequences.

- temporal difference learning with genetic algorithms
Temporal difference learning could provide a genetic algorithm with more accurate, less noisy fitness values than game outcome alone.

ideas to think about

- speed and smarts
What can we learn from Deep Blue about brute force? The class of brute force methods is a wide and important one.


updated 13 May 2012 - fixed the AutoClass links