need data mining
- too much data
- local lessons
- need fast ways to learn local truths
- need ways to tell when local truth is no longer useful

- need to talk humans and learners (so high elvel symbolic descriptions, please)
- need to be able to initialize the theory with expectations

(how am i going with top10 algorithms)

theory of everything

agents know what thye know, know what they don't know. know when then they go from know to dont know.

not just prediction. not "what is". but "what to do next". technically, not classifiers/predictors but contrast sets. differences between thing

repeated result: 
m <n vars control the remaining n-m vars
- so small teories are all you need

averaging over multiple approximations does better than generating one exact figure

learner1: knn
	knn
	log pre-processor
	baseline results

stats101
	mann whitney
	wilcoxon
	armstrong's rule.
	"the rig"

clusering:
	k-means
	genic
	idea (fastmap + tiles). good for visualization

discretiztion
	rbst
		node=<label,here,under,rows>
		here=stats
		under=stats
		stats=if numeric, then n,sum,sumSq else if discrete then uniques
		from= row ids
	equal width
	equal freq
	bore
	supervised
		variance-based
			min-ent
			min-var

learner3: simple bayes
	not naive

visualization
	bore
	nomograms
	b-squared
	
shrinkage
	infogain
	feature seletion
	instance selection	
		- data generation
			- with noise added
		- noise reduction
			- show with knn and nb
		- privacy

learner4: bam
	bam = var pruning +  neighborhood ranking (no bagging)

learner5: prism
	prism + range ordering

learner6: tar
	should run fast with the indexing of part 1

learner7:  which

anomaly detection

belief revision

multi-untility optimzation

importance sampling
	bagging, boosting, etc, all "importance sampling"