Intelligent Agent Foundations Forumsign up / log in
An implementation of modal UDT
post by Benja Fallenstein 662 days ago | Jessica Taylor, Luke Muehlhauser, Nate Soares and Patrick LaVictoire like this | discuss

One of the great advantages of working with Gödel-Löb provability logic is that it’s possible to implement an evaluator which efficiently checks whether a sentence in the language of GL is true. Mihaly and Marcello used this to write a program that checks whether two modal agents cooperate or defect against each other. Today, Nate and I extended this with an implementation of modal UDT, which allows us to check what UDT does on different decision problems—see Program.hs in the Github repository. No guarantees for correctness, since this was written rather quickly; if anybody is able to take the time to check the code, that would be very much appreciated!

The implementation of UDT is rather pleasing, I think. Here’s the informal definition of modal UDT, using PA + \(\ell\):

  • For every possible outcome \(j\), from best to worst:
    • For every possible action \(i\), in order:
      • If it’s provable in PA + \(\ell\) that “UDT takes action \(i\)” implies “the universe returns outcome \(j\)”, then take action \(i\).
  • If you’re still here, return a default action.

Here is the corresponding Haskell code:

udt :: (Enum a,Ord b,Show b,Enum b)
    => Int -> ModalProgram b a -> b -> ModalProgram b b
udt level univ dflt = modalProgram dflt $
  mFor $ \a ->
    mFor $ \b ->
      mIf (boxk level (Var b %> univ a)) (mReturn b)

Being able to write modal UDT like this makes it very easy to implement and try small variations on the code.

We used this code to check what modal UDT would do in a version of Newcomb’s problem where Omega uses proofs in PA (rather than simulations) to decide whether to put the money in the box; that is, it will put a million dollars in the first box if and only if it can prove that you will one-box. If our code is correct, it turns out that in this case, modal UDT will do whatever its default action was.

Earlier, we thought we had proved a different result on the whiteboard, but after the code disagreed with us, we went over it again and found a bug in our proof. After fixing that bug, we now have a manual proof that UDT will end up taking its default action in this scenario (which I’ll write about some other time). So looks like this can be a useful tool for figuring out this sort of thing!





Rewards and POMDP rather than
by Stuart Armstrong on Reward/value learning for reinforcement learning | 2 likes

What are the main differences
by Jessica Taylor on Reward/value learning for reinforcement learning | 0 likes

Nice! One thing that might be
by Patrick LaVictoire on (Non-)Interruptibility of Sarsa(λ) and Q-Learning | 0 likes

The sentence ending the first
by Abram Demski on Asymptotic Decision Theory | 0 likes

I agree with most of what you
by Wei Dai on Desiderata for decision theory | 0 likes

Well, the time to take a
by Vadim Kosoy on Updatelessness and Son of X | 0 likes

So my plan is to "solve" the
by Scott Garrabrant on Updatelessness and Son of X | 0 likes

I may be misunderstanding
by Wei Dai on Updatelessness and Son of X | 0 likes

>But we know that cooperation
by Wei Dai on Updatelessness and Son of X | 1 like

This does seem to be the
by Wei Dai on Updatelessness and Son of X | 3 likes

UDT, in its global policy
by Vladimir Nesov on Updatelessness and Son of X | 1 like

From my perspective, the
by Paul Christiano on A failed attempt at Updatelessness using Universal... | 0 likes

This is more or less what I
by Vadim Kosoy on Updatelessness and Son of X | 0 likes

There is a decent-sized
by Ryan Carey on Vector-Valued Reinforcement Learning | 0 likes

I still don't understand the
by Jessica Taylor on Counterfactual do-what-I-mean | 0 likes


Privacy & Terms (NEW 04/01/15)