Let’s reimplement EURISKO!

In the early 1980s Dou­glas Le­nat wrote EURISKO, a pro­gram Eliezer called “[maybe] the most so­phis­ti­cated self-im­prov­ing AI ever built”. The pro­gram re­port­edly had some high-pro­file suc­cesses in var­i­ous do­mains, like be­com­ing world cham­pion at a cer­tain wargame or de­sign­ing good in­te­grated cir­cuits.

De­spite re­quests Le­nat never re­leased the source code. You can down­load an in­tro­duc­tory pa­per: “Why AM and EURISKO ap­pear to work” [PDF]. Hon­estly, read­ing it leaves a pro­gram­mer still mys­tified about the in­ter­nal work­ings of the AI: for ex­am­ple, what does the main loop look like? Re­searchers sup­pos­edly an­swered such ques­tions in a more de­tailed pub­li­ca­tion, “EURISKO: A pro­gram that learns new heuris­tics and do­main con­cepts.” Ar­tifi­cial In­tel­li­gence (21): pp. 61-98. I couldn’t find that pa­per available for down­load any­where, and be­ing in Rus­sia I found it quite tricky to get a pa­per ver­sion. Maybe you Amer­i­cans will have bet­ter luck with your lo­cal library? And to the best of my knowl­edge no one ever suc­ceeded in (or even se­ri­ously tried) con­firm­ing Le­nat’s EURISKO re­sults.

To­day in 2009 this state of af­fairs looks laugh­able. A 30-year-old pivotal break­through in a large and im­por­tant field… that never even got re­pro­duced. What if it was a gi­gan­tic case of Clever Hans? How do you know? You’re sup­posed to be a sci­en­tist, lit­tle one.

So my pro­posal to the LessWrong com­mu­nity: let’s reim­ple­ment EURISKO!

We have some com­pe­tent pro­gram­mers here, don’t we? We have open source tools and lan­guages that weren’t around in 1980. We can build an open source im­ple­men­ta­tion available for all to play. In my book this counts as solid progress in the AI field.

Hell, I’d do it on my own if I had the god­damn pa­per.

Up­date: RichardKen­n­away has put Le­nat’s de­tailed pa­pers up on­line, see the com­ments.