Douglas Lenat’s program EURISKO is legendary in the AI community for a distinct real-world achievement: allowing Lenat to win the the Traveller TCS roleplaying game tournament two years in a row (and then semi-voluntarily not competing subsequent years). Lenat never released EURISKO’s source code, leaving how he managed to pull off this feat somewhat of a mystery. Yet Lenat’s later work based on EURISKO does not seem to have yielded anything else in the way of practical benefits.
My suspicion is that Lenat’s TCS win tells us more about TCS than about EURISKO, that TCS is likely a game that’s inherently vulnerable to the “find winning strategies by simulating a lots of games on a computer” meta-strategy. I’ve heard, for example, that battles are often tactically trivial, with the outcome of battles effectively determined by the composition of the two fleets (and fleet composition is what Lenat used EURISKO for). If that hypothesis is correct, though, it suggests it shouldn’t be necessary to reimplement EURISKO specifically to get a program that’s good at designing TCS fleets. If that turns out not to be the case, it would be evidence that there really is something special about EURISKO after all.
Does anyone know if anyone has tried this? As a novice computer programmer, I think it might be a good project to hone my programming skills. Input on how to approach such a project would be appreciated.
Replicating Douglas Lenat’s Traveller TCS win with publicly-known techniques
Douglas Lenat’s program EURISKO is legendary in the AI community for a distinct real-world achievement: allowing Lenat to win the the Traveller TCS roleplaying game tournament two years in a row (and then semi-voluntarily not competing subsequent years). Lenat never released EURISKO’s source code, leaving how he managed to pull off this feat somewhat of a mystery. Yet Lenat’s later work based on EURISKO does not seem to have yielded anything else in the way of practical benefits.
Some time ago on LessWrong, someone proposed trying to figure out what Lenat did and reimplementing EURISKO. But Eliezer is worried this could be dangerous. So I have another proposal: see if Lenat’s accomplishment can be replicated using machine learning and genetic programming techniques that are already publicly known.
My suspicion is that Lenat’s TCS win tells us more about TCS than about EURISKO, that TCS is likely a game that’s inherently vulnerable to the “find winning strategies by simulating a lots of games on a computer” meta-strategy. I’ve heard, for example, that battles are often tactically trivial, with the outcome of battles effectively determined by the composition of the two fleets (and fleet composition is what Lenat used EURISKO for). If that hypothesis is correct, though, it suggests it shouldn’t be necessary to reimplement EURISKO specifically to get a program that’s good at designing TCS fleets. If that turns out not to be the case, it would be evidence that there really is something special about EURISKO after all.
Does anyone know if anyone has tried this? As a novice computer programmer, I think it might be a good project to hone my programming skills. Input on how to approach such a project would be appreciated.