Ω 30

Crossposted from the AI Alignment Forum. May contain more technical jargon than usual.

This post explains a formal link between "what kinds of instrumental convergence exists?" and "what does VNM-coherence tell us about goal-directedness?". It turns out that VNM coherent preference orderings have the same statistical incentives as utility functions; most such orderings will incentivize power-seeking in the settings covered by the power-seeking theorems

In certain contexts, coherence theorems can have non-trivial implications, in that they provide Bayesian evidence about what the coherent agent will probably do. In the situations where the power-seeking theorems apply, coherent preferences do suggest some degree of goal-directedness. Somewhat more precisely, VNM-coherence is Bayesian evidence that the agent prefers to stay alive, keep its options open, etc.

However, VNM-coherence over action-observation histories tells you nothing about what behavior to expect from the coherent agent, because there is no instrumental convergence for generic utility functions over action-observation histories!

Intuition

The result follows because the VNM utility theorem lets you consider VNM-coherent preference orderings to be isomorphic to their induced utility functions (with equivalence up to positive affine transformation), and so these preference orderings will have the same generic incentives as the utility functions themselves.

Formalism

Let  be outcomes, in a sense which depends on the context; outcomes could be world-states, universe-histories, or one of several fruits. Outcome lotteries are probability distributions over outcomes, and can be represented as elements of the -dimensional probability simplex (ie as element-wise non-negative unit vectors).

A preference ordering  is a binary relation on lotteries; it need not be eg complete (defined for all pairs of lotteries). VNM-coherent preference orderings are those which obey the VNM axioms. By the VNM utility theorem, coherent preference orderings induce consistent utility functions over outcomes, and consistent utility functions conversely imply a coherent preference ordering.

Definition 1: Permuted preference ordering. Let  be an outcome permutation, and let  be a preference ordering.  is the preference ordering such that for any lotteries  if and only if .

EDIT: Thanks to Edouard Harris for pointing out that Definition 1 and Lemma 3 were originally incorrect.

Definition 2: Orbit of a preference ordering. Let  be any preference ordering. Its orbit  is the set

The orbits of coherent preference orderings are basically all the preference orderings induced by "relabeling" which outcomes are which. This is made clear by the following result:

Lemma 3: Permuting coherent preferences permutes the induced utility function. Let  be a VNM-coherent preference ordering which induces VNM-utility function , and let . Then  induces VNM-utility function , where  is any outcome.

Proof. Let  be any lotteries.

1. By the definition of a permuted preference ordering,  if and only if .
2. By the VNM utility theorem and the fact that  is coherent,  iff .
3. Since there are finitely many outcomes, we convert to vector representation: .
4. By associativity, .
5. But this is just equivalent to .

QED.

As a corollary, this lemma implies that if  is VNM-coherent, so is , since it induces a consistent utility function over outcomes.

Consider the orbit of any . By the VNM utility theorem, each preference ordering can be considered isomorphic to its induced utility function (with equivalence up to positive affine transformation).

Then let  be any utility function compatible with . By the above lemma, consider the natural bijection between the (preference ordering) orbit of  and the (utility function) orbit of , where .

When my theorems on power-seeking are applicable, some proportion of the right-hand side is guaranteed to make (formal) power-seeking optimal. But by the bijection and by the fact that the preference orderings incentivize the same things (by the VNM theorem in the reverse direction), the (preference ordering) orbit must have the exact same proportion of elements for which (lotteries representing formal) power-seeking are optimal.

Conversely, if we know that some set A of lotteries tends to be preferred over another set B of lotteries (in the preference order orbit sense), then the same argument shows that A tends to have greater expected utility than B (in the utility function orbit sense). This holds for all (utility function) orbits, because every utility function corresponds to a VNM-coherent preference ordering.

So: orbit-level instrumental convergence for utility functions is equivalent to orbit-level instrumental convergence for VNM-coherent preference orderings.

Implications

• Instrumental convergence does not exist when maximizing expected utility over action observation histories (AOH).
• Therefore, VNM-coherence over action observation history lotteries tells you nothing about what behavior to expect from the agent.
• Coherence over AOH tells you nothing because there is no instrumental convergence in that setting!
• In certain contexts, coherence theorems can have non-trivial implications, in that they provide Bayesian evidence about what the coherent agent will probably do.
• In the situations where the power-seeking theorems apply, coherent preferences do suggest some degree of goal-directedness.
• Somewhat more precisely, VNM-coherence is Bayesian evidence that the agent prefers to stay alive, keep its options open, etc.
• In some domains, preference specification may be more natural than utility function specification. However, in theory, coherent preferences and utility functions have the exact same statistical incentives.
• In practice, they will differ. For example, suppose we have a choice between specifying a reward function which is linear over state features, or of doing behavioral cloning on elicited human preferences over world states. These two methods will probably tend to produce different incentives.

The quest for better convergence theorems

Goal-directedness seems to more naturally arise from coherence over resources. (I think the word 'resources' is slightly imprecise here, because resources are only resources in the normal context of human life; money is useless when alone in Alpha Centauri, but time to live is not. So we want coherence over things-which-are-locally-resources, perhaps.)

In his review of Seeking Power is Often Convergently Instrumental in MDPs, John Wentworth wrote:

in a real-time strategy game, units and buildings and so forth can be created, destroyed, and generally moved around given sufficient time. Over long time scales, the main thing which matters to the world-state is resources - creating or destroying anything else costs resources. So, even though there's a high-dimensional game-world, it's mainly a few (low-dimensional) resource counts which impact the long term state space. Any agents hoping to control anything in the long term will therefore compete to control those few resources.

More generally: of all the many "nearby" variables an agent can control, only a handful (or summary) are relevant to anything "far away". Any "nearby" agents trying to control things "far away" will therefore compete to control the same handful of variables.

Main thing to notice: this intuition talks directly about a feature of the world - i.e. "far away" variables depending only on a handful of "nearby" variables. That, according to me, is the main feature which makes or breaks instrumental convergence in any given universe. We can talk about that feature entirely independent of agents or agency. Indeed, we could potentially use this intuition to derive agency, via some kind of coherence theorem; this notion of instrumental convergence is more fundamental than utility functions.

In his review of Coherent decisions imply consistent utilities, John wrote:

"resources" should be a derived notion rather than a fundamental one. My current best guess at a sketch: the agent should make decisions within multiple loosely-coupled contexts, with all the coupling via some low-dimensional summary information - and that summary information would be the "resources". (This is exactly the kind of setup which leads to instrumental convergence.) By making pareto-resource-efficient decisions in one context, the agent would leave itself maximum freedom in the other contexts. In some sense, the ultimate "resource" is the agent's action space. Then, resource trade-offs implicitly tell us how the agent is trading off its degree of control within each context, which we can interpret as something-like-utility.

This seems on-track to me. We now know what instrumental convergence looks like in unstructured environments, and how structural assumptions on utility functions affect the shape and strength of that instrumental convergence, and this post explains the precise link between "what kinds of instrumental convergence exists?" and "what does VNM-coherence tell us about goal-directedness?". I'd be excited to see what instrumental convergence looks like in more structured models.

Footnote representative: In terms of instrumental convergence, positive affine transformation never affects the optimality probability of different lottery sets. So for each (preference ordering) orbit element , it doesn't matter what representative we select from each equivalence class over induced utility functions — so we may as well pick !

Ω 30

Pingbacks
New Comment

Copying over a Slack comment from Abram Demski:

I think this post could be pretty important.

It offers a formal treatment of "goal-directedness" and its relationship to coherence theorems such as VNM, a topic which has seen some past controversy but which has -- till now -- been dealt with only quite informally. Personally I haven't known how to engage with the whole goal-directedness debate, and I think part of the reason for that is the vagueness of the idea. Goal-directedness doesn't seem that cruxy for most of my thinking, but some other people seem to really strongly perceive it as a crux for miri-type thought, and sometimes as a crux for AI risk more generally. (I once made a "tool AI" argument against AI risk myself, although in hindsight I would say that was all motivated cognition, which ignored the idea that even tool AI has to optimize strongly in order to have high capabilities.)

So, as I see it, there's been something of a stalemate between people who think the "goal-directed AI" vs "non-goal-directed AI" distinction is important for one reason or another, vs people who don't think that.

Alex Turner seems to give real technical meaning to this distinction, showing that most VNM-coherent preferences are indeed "goal directed" in the sense of acting broadly like we expect agents to act (that is, behaving in ways consistent with instrumental convergence). However, he also gives a class of VNM-coherent preferences which are not goal-directed in this sense, instead exhibiting essentially random behavior. This gives us a plausible formal proxy for the "goal directed vs not goal directed" distinction!

I'm not sure how it can/should carry the broader conversation forward, yet, but it seems like something to think about.

Thanks for writing this.

I have one point of confusion about some of the notation that's being used to prove Lemma 3. Apologies for the detail, but the mistake could very well be on my end so I want to make sure I lay out everything clearly.

First,  is being defined here as an outcome permutation. Presumably this means that 1)  for some ; and 2)  admits a unique inverse . That makes sense.

We also define lotteries over outcomes, presumably as, e.g., , where  is the probability of outcome . Of course we can interpret the  geometrically as mutually orthogonal unit vectors, so this lottery defines a point on the -simplex. So far, so good.

But the thing that's confusing me is what this implies for the definition of . Because  is defined as a permutation over outcomes (and not over probabilities of outcomes), we should expect this to be

The problem is that this seems to give a different EV from the lemma:

(Note that I'm using  as the dummy variable rather than , but the LHS above should correspond to line 2 of the proof.) Doing the same thing for the  lottery gives an analogous result. And then looking at the inequality that results suggests that lemma 3 should actually be " induces " as opposed to " induces ".

(As a concrete example, suppose we have a lottery  with the permutation . Then  and our EV is

Yet  which appears to contradict the lemma as stated.)

Note that even if this analysis is correct, it doesn't invalidate your main claim. You only really care about the existence of a bijection rather than what that bijection is — the fact that your outcome space is finite ensures that the proportion of orbit elements that incentivize power seeking remains the same either way. (It could have implications if you try to extend this to a metric space, though.)

Again, it's also possible I've just misunderstood something here — please let me know if that's the case!

Thanks! I think you're right. I think I actually should have defined  differently, because writing it out, it isn't what I want. Having written out a small example, intuitively,  should hold iff , which will also induce  as we want.

I'm not quite sure what the error was in the original proof of Lemma 3; I think it may be how I converted to and interpreted the vector representation. Probably it's more natural to represent  as , which makes your insight obvious.

The post is edited and the issues should now be fixed.