## AI ALIGNMENT FORUMAF

Of course, this interpretation requires a fair amount of reading between the lines, since the Jeffrey-Bolker axioms make no explicit mention of any probability distribution, but I don’t see any other reasonable way to interpret them,

Part of the point of the JB axioms is that probability is constructed together with utility in the representation theorem, in contrast to VNM, which constructs utility via the representation theorem, but takes probability as basic.

This makes Savage a better comparison point, since the Savage axioms are more similar to the VNM framework while also trying to construct probability and utility together with one representation theorem.

VNM does not start out with a prior, and allows any probability distribution over outcomes to be compared to any other, and Jeffrey-Bolker only allows comparison of probability distributions obtained by conditioning the prior on an event.

As a representation theorem, this makes VNM weaker and JB stronger: VNM requires stronger assumptions (it requires that the preference structure include information about all these probability-distribution comparisons), where JB only requires preference comparison of events which the agent sees as real possibilities. A similar remark can be made of Savage.

Starting with a prior that gets conditioned on events that correspond to the agent’s actions seems to build in evidential decision theory as an assumption, which makes me suspicious of it.

Right, that's fair. Although: James Joyce, the big CDT advocate, is quite the Jeffrey-Bolker fan! See Why We Still Need the Logic of Decision for his reasons.

I don’t think the motivation for this is quite the same as the motivation for pointless topology, which is designed to mimic classical topology in a way that Jeffrey-Bolker-style decision theory does not mimic VNM-style decision theory. [...] So a similar thing here would be to treat a utility function as a function from some lattice of subsets of (the Borel subsets, for instance) to the lattice of events.

Doesn't pointless topology allow for some distinctions which aren't meaningful in pointful topology, though? (I'm not really very familiar, I'm just going off of something I've heard.)

Isn't the approach you mention pretty close to JB? You're not modeling the VNM/Savage thing of arbitrary gambles; you're just assigning values (and probabilities) to events, like in JB.

Setting aside VNM and Savage and JB, and considering the most common approach in practice -- use the Kolmogorov axioms of probability, and treat utility as a random variable -- it seems like the pointless analogue would be close to what you say.

This can be resolved by defining worlds to be minimal non-zero elements of the completion of the Boolean algebra of events, rather than a minimal non-zero event.

Yeah. The question remains, though: should we think of utility as a function of these minimal elements of the completion? Or not? The computability issue I raise is, to me, suggestive of the negative.

# 39

This post has benefitted from discussion with Sam Eisenstat, Scott Garrabrant, Tsvi Benson-Tilsen, Daniel Demski, Daniel Kokotajlo, and Stuart Armstrong. It started out as a thought about Stuart Armstrong's research agenda.

In this post, I hope to say something about what it means for a rational agent to have preferences. The view I am putting forward is relatively new to me, but it is not very radical. It is, dare I say, a conservative view -- I hold close to Bayesian expected utility theory. However, my impression is that it differs greatly from common impressions of Bayesian expected utility theory.

I will argue against a particular view of expected utility theory -- a view which I'll call reductive utility. I do not recall seeing this view explicitly laid out and defended (except in in-person conversations). However, I expect at least a good chunk of the assumptions are commonly made.

# Reductive Utility

The core tenets of reductive utility are as follows:

• The sample space of a rational agent's beliefs is, more or less, the set of possible ways the world could be -- which is to say, the set of possible physical configurations of the universe. Hence, each world is one such configuration.
• The preferences of a rational agent are represented by a utility function from worlds to real numbers.
• Furthermore, the utility function should be a computable function of worlds.

Since I'm setting up the view which I'm knocking down, there is a risk I'm striking at a straw man. However, I think there are some good reasons to find the view appealing. The following subsections will expand on the three tenets, and attempt to provide some motivation for them.

If the three points seem obvious to you, you might just skip to the next section.

## Worlds Are Basically Physical

What I mean here resembles the standard physical-reductionist view. However, my emphasis is on certain features of this view:

• There is some "basic stuff" -- like like quarks or vibrating strings or what-have-you.
• What there is to know about the world is some set of statements about this basic stuff -- particle locations and momentums, or wave-form function values, or what-have-you.
• These special atomic statements should be logically independent from each other (though they may of course be probabilistically related), and together, fully determine the world.
• These should (more or less) be what beliefs are about, such that we can (more or less) talk about beliefs in terms of the sample space as being the set of worlds understood in this way.

This is the so-called "view from nowhere", as Thomas Nagel puts it.

I don't intend to construe this position as ruling out certain non-physical facts which we may have beliefs about. For example, we may believe indexical facts on top of the physical facts -- there might be (1) beliefs about the universe, and (2) beliefs about where we are in the universe. Exceptions like this violate an extreme reductive view, but are still close enough to count as reductive thinking for my purposes.

## Utility Is a Function of Worlds

So we've got the "basically physical" . Now we write down a utility function . In other words, utility is a random variable on our event space.

What's the big deal?

One thing this is saying is that preferences are a function of the world. Specifically, preferences need not only depend on what is observed. This is incompatible with standard RL in a way that matters.

But, in addition to saying that utility can depend on more than just observations, we are restricting utility to only depend on things that are in the world. After we consider all the information in , there cannot be any extra uncertainty about utility -- no extra "moral facts" which we may be uncertain of. If there are such moral facts, they have to be present somewhere in the universe (at least, derivable from facts about the universe).

One implication of this: if utility is about high-level entities, the utility function is responsible for deriving them from low-level stuff. For example, if the universe is made of quarks, but utility is a function of beauty, consciousness, and such, then needs to contain the beauty-detector and consciousness-detector and so on -- otherwise how can it compute utility given all the information about the world?

## Utility Is Computable

Finally, and most critically for the discussion here, should be a computable function.

To clarify what I mean by this: should have some sort of representation which allows us to feed it into a Turing machine -- let's say it's an infinite bit-string which assigns true or false to each of the "atomic sentences" which describe the world. should be a computable function; that is, there should be a Turing machine which takes a rational number and takes , prints a rational number within of , and halts. (In other words, we can compute to any desired degree of approximation.)

Why should be computable?

One argument is that should be computable because the agent has to be able to use it in computations. This perspective is especially appealing if you think of as a black-box function which you can only optimize through search. If you can't evaluate , how are you supposed to use it? If exists as an actual module somewhere in the brain, how is it supposed to be implemented? (If you don't think this sounds very convincing, great!)

Requiring to be computable may also seem easy. What is there to lose? Are there preference structures we really care about being able to represent, which are fundamentally not computable?

And what would it even mean for a computable agent to have non-computable preferences?

However, the computability requirement is more restrictive than it may seem.

There is a sort of continuity implied by computability: must not depend too much on "small" differences between worlds. The computation only accesses finitely many bits of before it halts. All the rest of the bits in must not make more than difference to the value of .

This means some seemingly simple utility functions are not computable.

As an example, consider the procrastination paradox. Your task is to push a button. You get 10 utility for pushing the button. You can push it any time you like. However, if you never press the button, you get -10. On any day, you are fine with putting the button-pressing off for one more day. Yet, if you put it off forever, you lose!

We can think of as a string like 000000100.., where the "1" is the day you push the button. To compute the utility, we might look for the "1", outputting 10 if we find it.

But what about the all-zero universe, 0000000...? The program must loop forever. We can't tell we're in the all-zero universe by examining any finite number of bits. You don't know whether you will eventually push the button. (Even if the universe also gives your source code, you can't necessarily tell from that -- the logical difficulty of determining this about yourself is, of course, the original point of the procrastination paradox.)

Hence, a preference structure like this is not computable, and is not allowed according to the reductive utility doctrine.

The advocate of reductive utility might take this as a victory. The procrastination paradox has been avoided, and other paradoxes with a similar structure. (The St. Petersburg Paradox is another example.)

On the other hand, if you think this is a legitimate preference structure, dealing with such 'problematic' preferences motivates abandonment of reductive utility.

# Subjective Utility: The Real Thing

We can strongly oppose all three points without leaving orthodox Bayesianism. Specifically, I'll sketch how the Jeffrey-Bolker axioms enable non-reductive utility. (The title of this section is a reference to Jeffrey's book Subjective Probability: The Real Thing.)

However, the real position I'm advocating is more grounded in logical induction rather than the Jeffrey-Bolker axioms; I'll sketch that version at the end.

## The View From Somewhere

The reductive-utility view approached things from the starting-point of the universe. Beliefs are for what is real, and what is real is basically physical.

The non-reductive view starts from the standpoint of the agent. Beliefs are for things you can think about. This doesn't rule out a physicalist approach. What it does do is give high-level objects like tables and chairs an equal footing with low-level objects like quarks: both are inferred from sensory experience by the agent.

Rather than assuming an underlying set of worlds, Jeffrey-Bolker assume only a set of events. For two events and , the conjunction exists, and the disjunction , and the negations and . However, unlike in the Kolmogorov axioms, these are not assumed to be intersection, union, and complement of an underlying set of worlds.

Let me emphasize that: we need not assume there are "worlds" at all.

In philosophy, this is called situation semantics -- an alternative to the more common possible-world semantics. In mathematics, it brings to mind pointless topology.

In the Jeffrey-Bolker treatment, a world is just a maximally specific event: an event which describes everything completely. But there is no requirement that maximally-specific events exist. Perhaps any event, no matter how detailed, can be further extended by specifying some yet-unmentioned stuff. (Indeed, the Jeffrey-Bolker axioms assume this! Although, Jeffrey does not seem philosophically committed to that assumption, from what I have read.)

Thus, there need not be any "view from nowhere" -- no semantic vantage point from which we see the whole universe.

This, of course, deprives us of the objects which utility was a function of, in the reductive view.

## Utility Is a Function of Events

The reductive-utility makes a distinction between utility -- the random variable itself -- and expected utility, which is the subjective estimate of the random variable which we use for making decisions.

The Jeffrey-Bolker framework does not make a distinction. Everything is a subjective preference evaluation.

A reductive-utility advocate sees the expected utility of an event as derived from the utility of the worlds within the event. They start by defining ; then, we define the expected utility of an event as -- or, more generally, the corresponding integral.

In the Jeffrey-Bolker framework, we instead define directly on events. These preferences are required to be coherent with breaking things up into sums, so = -- but we do not define one from the other.

We don't have to know how to evaluate entire worlds in order to evaluate events. All we have to know is how to evaluate events!

Jeffrey-Bolker doesn't say anything about computability. However, if we do want to address this sort of issue, it leaves us in a different position.

Because subjective expectation is primary, it is now more natural to require that the agent can evaluate events, without any requirement about a function on worlds. (Of course, we could do that in the Kolmogorov framework.)

Agents don't need to be able to compute the utility of a whole world. All they need to know is how to update expected utilities as they go along.

Of course, the subjective utility can't be just any way of updating as you go along. It needs to be coherent, in the sense of the Jeffrey-Bolker axioms. And, maintaining coherence can be very difficult. But it can be quite easy even in cases where the random-variable treatment of the utility function is not computable.

Let's go back to the procrastination example. In this case, to evaluate the expected utility of each action at a given time-step, the agent does not need to figure out whether it ever pushes the button. It just needs to have some probability, which it updates over time.

For example, an agent might initially assign probability to pressing the button at time , and to never pressing the button. Its probability that it would ever press the button, and thus its utility estimate, would decrease with each observed time-step in which it didn't press the button. (Of course, such an agent would press the button immediately.)

Of course, this "solution" doesn't touch on any of the tricky logical issues which the procrastination paradox was originally introduced to illustrate. This isn't meant as a solution to the procrastination paradox -- only as an illustration of how to coherently update discontinuous preferences. This simple is uncomputable by the definition of the previous section.

It also doesn't address computational tractability in a very real way, since if the prior is very complicated, computing the subjective expectations can get extremely difficult.

We can come closer to addressing logical issues and computational tractability by considering things in a logical induction framework.

# Utility Is Not a Function

In a logical induction (LI) framework, the central idea becomes "update your subjective expectations in any way you like, so long as those expectations aren't (too easily) exploitable to Dutch-book." This clarifies what it means for the updates to be "coherent" -- it is somewhat more elegant than saying "... any way you like, so long as they follow the Jeffrey-Bolker axioms."

This replaces the idea of "utility function" entirely -- there isn't any need for a function any more, just a logically-uncertain-variable (LUV, in the terminology from the LI paper).

Actually, there are different ways one might want to set things up. I hope to get more technical in a later post. For now, here's some bullet points:

• In the simple procrastination-paradox example, you push the button if you have any uncertainty at all. So things are not that interesting.
• In more complicated examples -- where there is some real benefit to procrastinating -- a LI-based agent could totally procrastinate forever. This is because LI doesn't give any guarantee about converging to correct beliefs for uncomputable propositions like whether Turing machines halt or whether people stop procrastinating.
• Believing you'll stop procrastinating even though you won't is perfectly coherent -- in the same way that believing in nonstandard numbers is perfectly logically consistent. Putting ourselves in the shoes of such an agent, this just means we've examined our own decision-making to the best of our ability, and have put significant probability on "we don't procrastinate forever". This kind of reasoning is necessarily fallible.
• Yet, if a system we built were to do this, we might have strong objections. So, this can count as an alignment problem. How can we give feedback to a system to avoid this kind of mistake? I hope to work on this question in future posts.