Saturday, October 20, 2007

Occam's Meta-Razor

Let me define the Occam's Meta-Razor Problem as follows: What is the smallest and simplest set of basic philosophical postulates that a rational agent needs in order to act in a way that is intuitively satisfactory? The goal is that the behavior should satisfice, even if it's not necessarily optimal. Call this the Occam's Meta-Razor problem.

Intuitively, I think we want three items:

1. A simple way to analyze probabilities. Something like Solomonoff Induction might satisfice, if the Pascal's Mugging problem were solved.

2. A utility function. An initial start might be, “Maximize the expected amount of X in the Universe,” where X is some weighted combination of happiness, freedom from pain, autonomy, etc. A satisfactory but simple description for X would be difficult to unambiguously specify, especially in the case where the agent wields super-human intelligence. Two of many possible pitfalls:

  • For almost all X, the current set of humans who are alive (and humanity in general) are going to be sub-optimal, from the point-of-view of the agent. However, we want the agent to decide against wiping out humanity and replacing it with species that are “more worthy” according to its utility function.
  • We would want some portion of X to include the concept of “autonomy” and preserve our abilities to make informed, uncoerced decisions. But, a sufficiently smart agent could peacefully convince (trick?) me into making any number of ludicrous decisions. It's not clear how to unambiguously define “coercion” in the case of a super-intelligent agent.

3. A simple decision theory, such as Evidential Decision Theory (which I believe subsumes Hofstadter superrationality), or Causal Decision Theory (which is the standard in mainstream Game Theory.) Either should satisfice, though I regard Evidential Decision Theory as much simpler.

Being philosophical principles, obviously these can't be directly used to create a real, resource-limited AGI; for example, Solomonoff Induction is too slow for practical use.

But, as a set of normative philosophical principals for a human being to use, these seem like a reasonable starting point.


[edit -- decided to call it "Occam's Meta-Razor" rather than "Meta-Occam's Razor"]

No comments: