vetta project

In my opinion, the biggest theoretical problem with universal intelligence is the dependence on the reference UTM. The invariance property of Kolmogorov complexity provides some protection, but it’s not really enough. Years ago I thought about a number of ways to try to remedy this. I don’t recall all the possibilities that I considered, however I do remember that I never found one that I was completely happy with.

Talking to Peter Dayan about this the other day, he suggested the following: run the test multiple times always sampling from the same reference machine and allowing the agent to keep state across trials. Over time a very smart agent will start to learn the environmental sampling distribution and thus will be able to overcome not knowing the reference machine.

I remember talking to Marcus Hutter about measuring the intelligence of an agent by enumerating from the simplest environment up. The idea being that the simplest will have the biggest impact on the universal intelligence result anyway. Of course this would open up the possibility that the agent could work out the enumeration method and then not have to “learn” about a new environment by having to interact with it. That would be cheating. If we allow the agent to learn across trials but randomise which environments it sees each time, that should allow smart agents to learn to overcome the compiler constant and yet still not allow it to cheat.

It should then be possible to prove that an AIXI agent has maximal universal intelligence even when the agent and the test use difference reference machines.