A very non-technical explanation of the basics of infra-Bayesianism

8Vanessa Kosoy

6matolcsid

9Vanessa Kosoy

5Chris_Leong

2Søren Elverlin

6matolcsid

2Lao Mein

4matolcsid

1martinkunev

New Comment

Some quick comments:

- There is a reason why we
*shouldn't*expect agents to always guess correctly the bits divisible by 4: it might be that guessing wrong causes some*other*bits to become more easily guessable. Moreover, the "causality" might be of any nature (e.g. physical or pure "logical"). - For a summary of the results of IB, see this.
- The claim that infra-Bayesian physicalism hasn't produced any results so far is not fair. It hasn't produced
*major*results, but the original article does contain some theorems.

Okay, maybe I was somewhat unfair in saying there are no results. Sill, I think it's good to distinguish "internal results" and "external results". Take the example of complex analysis: we have many beautiful results about complex holomorphic functions, like Cauchy's integral formula. I call these internal results. But what made complex analysis so widely studied is that it could be used to produce some external results, like calculating the integral under the bell curve or proving the prime number theorem. These are questions that interested people even before holomorphic functions were invented, so proving them gave a legitimacy to the new complex analysis toolkit. Obviously, Cauchy's integral formula and the like are very useful too, as we couldn't reach the external results without understanding the toolkit itself better with the internal results. But my impression is that John was asking for an explanation of the external results, as they are more of an interest in an introductory post.

I count the work on Newcomb as an external result: "What learning process can lead to successfully learning to one-box in Newcomb's game?" is a natural question someone might ask without hearing about infra-Bayesianism, and I think IB gives a relatively natural framework for that (although I haven't looked into this deeply, and I don't know exactly how natural or hacky it is). On the other hand, from the linked results, I think the 1st, 4th and 5th are definitely internal results, I don't understand so can't comment of the 3rd, and the 2nd is Newcomb which I acknowledge. Similarly, I think IBP itself tries to answer an external question (formalizing naturalized induction), but I'm not convinced it succeeds in that, and I think the theorems are mostly internal results, and not something I would count as an external evidence. (I know less about this, so maybe I'm missing something).

In general, I don't deny IB has many internal results, which I acknowledge to be a necessary first step. But I think that John was looking for external results here, and in general my impression is that people seem to believe that there are more external results than there really are (did I mention the time I got a message from a group of young researchers asking if I thought "if it is currently feasible integrating multiple competing scientific theories into a single infra-Bayesian model"?) So I think it' useful to be more clear about that we don't have that many external results.

I partially agree, but the distinction between "internal" and "external" results is more fuzzy and complicated than you imply. Ultimately, it depends on the original problem you started with. For example, if you only care about prime numbers, then most results of complex analysis are "internal", with the exception of results that imply something about the distribution of prime numbers. However, if complex functions are a natural way to formalize the original problem, then the same results become "external".

In our case, the original problem is "creating a mathematical theory of intelligent agents". (Or rather, the problem is "solving AI alignment", or "preventing existential risk from AI", or "creating a flourishing future for human civilization", but let's suppose that the path from there to "creating a mathematical theory of intelligent agents" is already clear; in any case that's not related specifically to IB.) Infra-Bayesianism is supposed to be an actual *ingredient* in this theory of agents, not just some tool brought from the outside. In this sense, it already starts out as somewhat "external".

To give a concrete example, you said that results about IB multi-armed bandits are "internal". While I agree that these results are only useful as very simplistic toy models, they are potentially necessary steps towards stronger regret bounds in the future. At what point does it become "external"? Taking it to the extreme, I can imagine regret bounds so powerful, that they would serve as substantial evidence that an algorithm satisfying them is AGI or close to AGI. Would such a result still be "internal"?! Arguably not, because AGI algorithms are very pertinent to what we're interested in!

You can also take the position that any result without direct applications to existing, practical, economically competitive AI systems is "internal". In such case, I am comfortable with a research programme that only has "internal" results for a long time (although not everyone would agree). But this also doesn't seem to be your position, since you view results about Newcombian problems as "external".

How is it less of a hack if we are using measures instead of probability distributions? Also, how is "all losses are wiped out" less contrived than infinite utility?

We discussed this post in the AISafety.com Reading Group, and have a few questions about it and infra-bayesianism:

- The image on top of the sequence on Infra-Bayesianism shows a tree, which we interpret as a game-tree, with Murphy and an agent alternating in taking actions. Can we say anything about such a tree? E.g. Complexity, Pruning, etc?
- There was some discussion about if an infra-bayesian agent could be Dutch-booked. Is this possible?
- Your introduction makes no attempt to explain "convexity", which seems like a central part of Infra-Bayesianism. If it is central, what would be a good one-paragraph summary?
- Will any sufficiently smart agent be infra-bayesian? To be precise, can you replace "Bayesian" with "Infra-Bayesian" in this article: https://arbital.com/p/optimized_agent_appears_coherent/ ?

- No idea. I don't think it's computationally very tractable. If I understand correctly, l Vanessa hopes there will be computationally feasible approximations, but there wasn't much research into computational complexity yet, because there are more basic unsolved questions.
- I'm pretty sure that no. An IB agent (with enough compute) plans for the long run and doesn't go into a chain of deals that leaves it worse of than not doing anything. In general, IB solves the "not exactly Bayesian expected utility maximizer but still can't be Dutch booked problem" by potentially refusing to take either side of a bet: if it has Knightian uncertainty about whether a probability is lower or higher than 50%, it will refuse to bet at even odds either for or against. This is something that humans actually often do, and I agree with Vanessa that a decision theory can be allowed to do that.
- I had a paragraph about it:

"Here is where convex sets come in: The law constrains Murphy to choose the probability distribution of outcomes from a certain set in the space of probability distributions. Whatever the loss function is, the worst probability distribution Murphy can choose from the set is the same as if he could choose from the convex hull of the set. So we might as well start by saying that the law must be constraining Murphy to a convex set of probability distributions."

As far as I can tell, this is the reason behind considering convex sets. This makes convexity pretty central: laws are very central, and now we are assuming that every law is a convex set in the space of probability distributions. - Vanessa said that her guess is yes. In the terms of the linked Arbital article, IB is intended to be an example of "There could be some superior alternative to probability theory and decision theory that is Bayesian-incoherent". Personally, I don't know, I think that the article's "A cognitively powerful agent might not be sufficiently optimized" possibility feels more likely in the current paradigm, I can absolutely imagine the first AIs to become a world-ending threat not being very coherent. Also, IB is just an ideal, real-world smart agents will be at best approximations of infra-Bayesian agents (same holds for Bayesianism). Vanessa's guess is that understanding IB better will still give us useful insights into these real-world models if we view them as IB approximations, I'm pretty doubtful, but maybe. Also, I feel that the problem I write about in my post on the monotonicity principle points at some deeper problem in IB which makes me doubtful whether sufficiently optimized agents will actually use (approximations of) the minimax thinking prescribed by IB.

Regarding 4: given that infra-Bayesianism is maximally paranoid, shouldn't it have lower performance relative to decision-making theories like regular Bayes under many non-adversarial conditions? If the training set does not contain many instances of adversarial information, then shouldn't we expect agents to adopt Bayes instead of infra-Bayes?

I think Vanessa would argue that "Bayesianism" is not really an option. The non-realizability problem in Bayesianism is not just some weird special case, but the normal state of things: Bayesianism assumes that we have hypotheses fully describing the world, which we very definitely don't have in real life. IB tries to be less demanding, and the laws in the agent's hypothesis class don't necessarily need to be that detailed. I am relatively skeptical of this, and I believe that for an IB agent to work well, the laws in its hypothesis class probably also need to be unfeasibly detailed. So both "adopting Bayes" and "adopting infra-Bayes" fully is impossible. We probably won't have such a nice mathematical model for the messy decision process a superintelligence actually adopts, the question is whether thinking about it as an approximation of Bayes or infra-Bayes gives us a more clear picture. It's a hard question, and IB has an advantage in that the laws need to be less detailed, and a disadvantage that I think you are right about it being unnecessarily paranoid. My personal guess is that nothing besides the basic insight of Bayesianism ("the agent seems to update on evidence, sort of following Bayes-rule") will be actually useful in understanding the way an AI will think.

"the agent guesses the next bits randomly. It observes that it sometimes succeeds, something that wouldn't happen if Murphy was totally unconstrained"

Do we assume Murphy knows how the random numbers are generated? What justifies this?

IntroductionAs a response to John Wentworth's public request, I try to explain the basic structure of infra-Bayesian decision-making in a nutshell. Be warned that I significantly simplify some things, but I hope it gives roughly the right picture.

This post is mostly an abridged version of my previous post Performance guarantees in in classical learning and infra-Bayesianism. If you are interested in the more detailed and less sloppy version, you can read it there, it's a little more technical, but still accessible without serious background knowledge.

I also wrote up my general thoughts and criticism on infra-Bayesianism, and a shorter post explaining how infra-Bayesianism leads to the monotonicity principle.

Classical learning theoryInfra-Bayesianism was created to address some weak spots in classical learning theory. Thus, we must start by briefly talking about learning theory in general.

In classical learning theory, the agent has a hypothesis class, each hypothesis giving a full description of the environment. The agent interacts with the environment for a long time, slowly accumulating loss corresponding to certain events. The agent has a time discount rate γ, and its overall life-time loss is calculated by weighing the losses it receives through its history by this time discount.

The regret of an agent in environment e is the difference between the loss the agent actually receives through its life, and the loss it would receive if it followed the policy that is optimal for the environment e.

We say that a policy successfully learns a hypothesis class H, if it has low expected regret with respect to every environment e described in the hypothesis class.

How can a policy achieve this? In the beginning, the agent takes exploratory steps and observes the environment. If the observations are much more likely in environment e1 than in environment e2, then the agent starts acting in ways that make more sense in environment e1, and starts paying less attention to what would be a good choice in e2. This works well, because it doesn't cause much expected regret in e2 if the policy starts making bad choices in situations that are very unlikely to occur in e2: by definition, this only rarely causes problems.

Bayesian updating is a special case of this: originally, there is a prior on every hypothesis, then it's updated by how unlikely the observations are according to the particular hypothesis.

But this is just one possible way of learning. Assume that the hypothesis class H has certain favorable properties (it excludes traps), and the discount rate γ is close to 1, that is, the agent can be patient and devote a long time to exploration. Then, there are many algorithms that do a reasonably good job at learning: whichever e∈H environment the agent is placed at, it will have low expected regret compared to the best possible policy in e.

The non-realizability problemSo we selected a policy that has low expected regret in every environment in the hypothesis class. Unfortunately, we still don't know anything about how this policy does in environments that are not in the hypothesis class (technical term: the environment is non-realizable). This is a big problem, because it's certain that in real life, the true environment will not be in the agent's hypothesis class. First, because it's just practically impossible to have a full model of the world, down to he atomic level. Second, the environment includes the agent itself, which makes his theoretically impossible too.

Let's look at a simple example. The agent observes a sequence of 0s and 1s, and is tasked with predicting the next token (it gets 0 loss for guessing correctly, and 1 loss for incorrectly, and it has a shallow discount rate, which means it has long time for learning).

The agent has only a limited hypothesis class, it only contains hypotheses that claim that the observed sequence is produced by a finite-state machine (with an output function and with a potentially probabilistic transition function). It even selected a policy that has low regret in all these environments.

Meanwhile, the actual sequence it observes is 1 for square-free indices and 0 otherwise. Do we have any guarantee that the agent will do reasonably well in this environment?

Unfortunately, not. For every reasonable policy, one can attach to it this extra condition: "If you observed the square-free sequence so far, then guess 0 on square-free indices and 1 otherwise". It's the exact opposite of what it should be doing.

This is really dumb. Maybe we shouldn't expect it to do a perfect job in predicting the sequence, but in the observed sequence so far every fourth bit was 0, and the agent is observing this pattern for millions of bits, surely it should pick up on that and guess at least the fourth bits correctly. Or at least do something equally good as predicting the fourth bits, instead of just completely failing.

On the other hand, this additional failure mode condition wouldn't make the policy significantly worse in any of the environments in its hypothesis class. No finite state machine produces the square-free sequence for too long, so the policy having this failure mode on the square-free sequence doesn't cause significant expected regret for any sequence in the hypothesis class (assuming the time discount γ→1.)

Why would a policy have a failure mode like that? Well, there is no clear reason for that, and if there is a failure in real life, it probably won't look like "the policy almost always behaves reasonably, but goes crazy on one particular sequence, the square-free one". But this example is intended to show that theoretically it can happen, we have no performance guarantee on environments outside the hypothesis class. So this simplistic toy example demonstrates that a policy having low regret on a relatively broad hypothesis class (like sequences produced by finite state automatons) doesn't guarantee that it won't behave in really stupid ways in an environment outside the hypothesis class (like keep guessing always wrong, even though at least learning that the fourth bits are 0 should be really simple). In my longer post, I give some more examples and define the terms more precisely.

At this point, we might ask how likely it is for a policy selected to do well on a class of environments to do very stupid things in the environment we actually deploy it in. This depends on the specific way we select the policy, what the hypothesis class and the real environment exactly are, and what we mean by "very stupid things".

But Vanessa's agenda is more ambitious and more theoretical than that, and asks whether we can construct a new learning theory, in which a not-extremely-broad hypothesis class can be enough to really guarantee that the agent doesn't do very stupid things in any environment.

Infra-BayesianismAn infra-Bayesian learning agent assumes that the environment is controlled by a malevolent deity, Murphy. Murphy wants the agent to receive the maximum possible life-time loss (calculated with the time discount γ), and Murphy can read the agent's mind and knows its full policy from the beginning.

However, Murphy's actions are constrained by some law. The agent originally doesn't know what this law is, that's what it needs to learn over the course of its life. Accordingly, the agent's hypothesis class doesn't range over descriptions of "what the world is like" but over laws "how Murphy is constrained".

The agent selects a policy that has low regret with respect to every law Θ in its hypothesis class. Regret is defined as the difference between the actual loss received and the loss that the optimal policy chosen for Θ would receive against Murphy if Murphy was constrained only by the law Θ. There are theorems showing that for some pretty broad hypothesis classes of laws we can select a policy that has low regret with respect to all of them (as γ→1).

It's useful to note here that the laws can also be probabilistic in nature. For example: "On the seventh digits, Murphy can choose the probability of 0 being displayed between 17% and 39%. Then the actual bit is randomly generated based on the probability Murphy chose."

Here is where convex sets come in: The law constrains Murphy to choose the probability distribution of outcomes from a certain set in the space of probability distributions. Whatever the loss function is, the worst probability distribution Murphy can choose from the set is the same as he could choose from the convex hull of the set. So we might as well start by saying that the law must be constraining Murphy to a convex set of probability distributions.

Go back to our previous example of an agent doing sequence-prediction of the square-free sequence. The agent has this relatively narrow hypothesis class: "There is a finite-state machine (with a potentially probabilistic transition function) whose states are assigned to 0,1 or X. If the state is assigned to is 0 or 1, then Murphy is constrained to show that as the next bit of the sequence. If the state is assigned to X, Murphy can choose the next bit." There is a theorem that shows that there can be policies that have low regret with respect to all these hypotheses (as γ→1). Assume the agent has a policy like that.

Initially, the agent guesses the next bits randomly. It observes that it sometimes succeeds, something that wouldn't happen if Murphy was totally unconstrained. So it can rule out this hypothesis. As time goes on, it observes that every fourth bit is 0, so it tries guessing 0 on fourth bits... and it gets it right! Murphy knows the agent's full policy, so the fact that the agent can successfully guess the fourth digits suggests that Murphy is constrained on the fourth bits to output 0. This way, the agent starts to pay more attention to hypotheses that correctly predict real patterns in the sequence (like every fourth and ninth digit being 0), and starts to act accordingly (by guessing 0 on these bits). It's important to remember though that Murphy plans for the long-run and can read the agent's full policy, so the agent needs to keep an eye out for the possibility that Murphy is not actually constrained in this way, just wants to teach the agent the wrong lesson as part of a long-term plot. So the agent might need to revise this hypothesis when things start to go wrong, but until the 0s keep relibly coming on the fourth digits, it makes sense to guess that.

Can the same problem arise as before? Namely, that the agent just has a failure mode on observing the square-free sequence, which just makes it act stupidly.

Well, it depends on how stupidly. It's still not guaranteed that the agent successfully "learns" the pattern to always guess 0 on the fourth digits. In fact, we can prove that it's still possible that it guesses 1 for every digit if it observes the square-free sequence.

But always guessing 1 would at least mean getting lots of digits right (even if not the fourth digits). That's something we actually have a performance guarantee for: if an infra-Bayesian agent observes a sequence where every fourth digit is 0, then it is guaranteed to get at lest 14 of the digits right.

Otherwise, if observing the square-free sequence produced such a serious failure mode that would make the agent guess less than 14 of the digits right, then if Murphy was only constrained by the law "every fourth digit must be 0", then Murphy would intentionally produce the square-fee sequence to induce the failure mode. That would mean that a policy with a failure mode like that would have a non-negligible regret with respect to the law "every fourth digit must be 0", as it does significantly worse than the optimal policy of always guessing 0 on the fourth digits. However, we assumed that the policy has low regret with respect to every law in its hypothesis class, and this simple law is included in that. Contradiction.

Thus, we proved that an infra-Bayesian learning agent with a relatively narrow hypothesis class will always guess at least 14 of the digits right if it observes a sequence in which every fourth digit is 0. I'm only moderately satisfied with this result, it would have felt more natural to prove that it will learn to get the actual fourth digits right, but that's not true. Still, even this performance guarantee is better than what we get from classical learning theory.

(Also, Vanessa has a different intuition here and doesn't think that correctly predicting the fourth bits would have been a more natural goal than just getting 14 of the digits right somewhere. "Rationality is systematized winning: Ultimately, what's important is how overall successful the agent is at achieving its goals, and the particular actions the agent takes are instrumental." That's also fair point.)

The results of infra-BayesianismJohn's request was to explain "the major high-level definitions and results in infra-Bayes to date". I really don't know of many "major high-level results".

To the best of my knowledge, the main motivation behind the development of infra-Bayesianism was that it helps with the non-realizability problem, that is, the above-described example generalizes to other situations, and we can prove interesting performance guarantees in general environments if we just assume that the infra-Bayesian agent has low regret on a relatively narrow hypothesis class. I don't know of strong results about this yet, and I'm relatively skeptical whether we will get further with the research of performance guarantees. You can read more details in my longer, technical post.

Newcomb's problem(lower confidence, this was not my research direction)In Newcomb's problem, it's just really convenient that we are already assuming that the environment is controlled by Murphy

who knows the agent's full policy from the beginning, effectively reading its mind.We can include some laws in the hypothesis class that say that if Murphy doesn't fill box A but the agent still one-boxes in that turn, then all the agent's losses are wiped out forever. I feel that including this is kind of hacky, but it feels less hacky in the more general framework where we are talking about measures instead of probability distributions, and a law like this can be relatively naturally encoded in that framework.

If the agent has some laws like that in its hypothesis class, then it will do some exploratory steps where it just one-boxes to see what happens.

If there is no correlation between one-boxing and finding treasure (as it's usually the case in our world), it soon learns not to leave money-boxes unopened and does exploratory one-boxing just very very occasionally.

On the other hand, if it's actually playing against Omega in Newcomb's game, then it will soon realize that the law in its hypothesis class that best explains the observations is the one according to which every time it decides to one-box, Murphy fills the box with treasure, otherwise all the loss would be wiped out which Murphy really wants to avoid.

I'm unsure how I feel about this solution to Newcomb's problem, but I think I agree with the main point: you are not born knowing how the game works and what the good action is, you learn it from playing the game yourself or watching others play. And infra-Bayesianism gives a relatively natural framework for learning the right lesson that you should one-box.

If you want to learn more about this, I recommend Thomas Larsen's relatively beginner-friendly explanation. Note that his post still relies on the older framework in which agents had utilities an not losses, so "all the loss is wiped out" is equivalent to "the agent receives infinite utility". This made the previous formalism somewhat more contrived than it currently is.

Other resultsI really don't know about other results building on infra-Bayesianism. Infra-Bayesian logic and Infra-Bayesian Physicalism exist, but they feel like mostly separate fields that only borrow some mathematical formalism from the original infra-Bayes, and they can't be easily explained based on the the non-formalized version of infra-Bayes I presented above. Also, as far as I know, they didn't really produce any actual results yet.

Maybe the other mentees who looked into other parts of infra-Bayesianism can link results in the comments if they find something I missed.