Intelligent Agent Foundations Forumsign up / log in
A permutation argument for comparing utility functions
discussion post by Stuart Armstrong 31 days ago | 2 comments

When doing intertheoretic utility comparisons, there is one clear and easy case: when everything is exactly symmetric.

This happens when, for instance, there exists \(u\) and \(v\) such that \(p([u])=p([v])=0.5\) and there exists a map \(\sigma : \mathbb{S}\to\mathbb{S}\), such that \(\sigma^2\) is the identity (hence \(\sigma\) is an involution) and for all \(s\in\mathbb{S}\)), \(u(s)=v(\sigma(s))\).

Note that this implies that \((u(s),v(s))=(v(\sigma(s)),u(\sigma(s))\), so \(\sigma\) is essentially a ‘reflection’.

Then, since everything is so symmetric, we can say there is no way of distinguishing \(u\) from \(v\), so the correct approach is to maximise \([u+v]\).

See the following graph, with the strategy to be followed marked in red:


Symmetry is good as far as it goes, but is very fragile. It doesn’t say anything about what happens when \(p([u])=49/100\) and \(p([v])=51/100\), for instance.

There is an argument, however, that resolves the unequal probability cases and extends the results to non-symmetric cases.

Consider for instance the case where \(u\) and \(v\) are as follows, for \(5\) strategies in \(\mathbb{S}\):

Nothing obvious springs to mind as to what the best normalisation process is – the setup is clearly unsymmetrical, and four of the five options are on the Pareto boundary. But let’s rescale and translate the utilities:

This is still unsymmetrical, but note that \(u\) and \(v\) have the same values: \(-1\), \(0.5\), \(2\), \(3\), and \(3.5\).

Thus there is a permutation \(\rho:\mathbb{S}\to\mathbb{S}\) such that for all \(s\in\mathbb{S}\), \(u(s)=v(\rho(s))\).

Another type of uncertainty

This permutation \(\rho\) allows us to transform the uncertainty about one’s own values (which we don’t know how to handle) to other types of uncertainty (which we can).

How so? Let \(\mathbb{S}'\) be another copy of \(\mathbb{S}\), but with different labels, and let \(i:\mathbb{S}\to\mathbb{S}'\) be the identity map that re-assigns each element to its original label.

Then instead of seeing \(u\) and \(v\) as different utility functions on \(\mathbb{S}\), we can see them as both being the same utility function \(w\) on \(\mathbb{S}'\), with uncertainty over a map \(m\) from \(\mathbb{S}\) to \(\mathbb{S}'\). This map \(m\) is \(i\) with probability \(p([u])\) and \(i\circ\rho\) with probability \(p([v])\).

Thus the agent should see itself as a \(w\)-maximiser, with standard uncertainty over the map \(m\) (this could be seen as uncertainty over the consequences of choosing a strategy). This means it will maximise \([p([u])u+p([v])v]\), as long as \(u\) and \(v\) are related by a permutation on \(\mathbb{S}\).

In this particular case, if \(p([u])\) and \(p([v])\) are close to \(0.5\), this means that the red strategy will be selected:

Note that permutations includes the symmetric case where \(\rho\) is an involution, so the symmetric argument now extends to cases where \(p([u])\neq p([v])\).


For permutations, this argument claims that the correct approach is to use Individual normalisations. Indeed, every normalisation presented here would reach the same result.

This doesn’t prove that individual normalisation is necessarily correct – you can imagine a general system that only give individual normalisations on permutation problems – but is suggestive none the less.

by Owen Cotton-Barratt 31 days ago | link

I’m not sure I’ve fully followed, but I’m suspicious that you seem to be getting something for nothing in your shift from a type of uncertainty that we don’t know how to handle to a type we do.

It seems to me like you must be making an implicit assumption somewhere. My guess is that this is where you used \(i\) to pair \(\mathbb{S}\) with \(\mathbb{S}'\). If you’d instead chosen \(j = i \circ \rho\) as the matching then you’d have uncertainty between whether \(m\) should be \(j\) or \(\rho^{-1}\circ j\). My guess is that generically this gives different recommendations from your approach.


by Stuart Armstrong 31 days ago | Owen Cotton-Barratt likes this | link

Nope! That gives the same recommendation (as does the same thing if you pre-compose with any other permutation of \(\mathbb{S}\)). I thought about putting that fact in, but it took up space.

The recommendation given in both cases is just to normalise each utility function individually, using any of the methods that we know (which will always produce equivalent utility classes in this situation).






The "benign induction
by David Krueger on Why I am not currently working on the AAMLS agenda | 0 likes

This comment is to explain
by Alex Mennen on Formal Open Problem in Decision Theory | 0 likes

Thanks for writing this -- I
by Daniel Dewey on AI safety: three human problems and one AI issue | 1 like

I think it does do the double
by Stuart Armstrong on Acausal trade: double decrease | 0 likes

>but the agent incorrectly
by Stuart Armstrong on CIRL Wireheading | 0 likes

I think the double decrease
by Owen Cotton-Barratt on Acausal trade: double decrease | 0 likes

The problem is that our
by Scott Garrabrant on Cooperative Oracles: Nonexploited Bargaining | 1 like

Yeah. The original generator
by Scott Garrabrant on Cooperative Oracles: Nonexploited Bargaining | 0 likes

I don't see how it would. The
by Scott Garrabrant on Cooperative Oracles: Nonexploited Bargaining | 1 like

Does this generalise to
by Stuart Armstrong on Cooperative Oracles: Nonexploited Bargaining | 0 likes

>Every point in this set is a
by Stuart Armstrong on Cooperative Oracles: Nonexploited Bargaining | 0 likes

This seems a proper version
by Stuart Armstrong on Cooperative Oracles: Nonexploited Bargaining | 0 likes

This doesn't seem to me to
by Stuart Armstrong on Change utility, reduce extortion | 0 likes

[_Regret Theory with General
by Abram Demski on Generalizing Foundations of Decision Theory II | 0 likes

It's not clear whether we
by Paul Christiano on Infinite ethics comparisons | 1 like


Privacy & Terms