# Attainable Utility Landscape: How The World Is Changed

(This is one interpretation of the prompt, in which you haven’t *chosen* to go to the moon. If you imagined yourself as more prepared, that’s also fine.)

If you were plopped onto the moon, you’d die pretty fast. Maybe the “die as quickly as possible” AU is high, but not much else—not even the “live on the moon” AU! We haven’t yet reshaped the AU landscape on the moon to be hospitable to a wide range of goals. Earth is special like that.

## AU landscape as a unifying frame

Attainable utilities are calculated by winding your way through possibility-space, considering and discarding possibility after possibility to find the best plan you can. This frame is unifying.

Sometimes you advantage one AU at the cost of another, moving through the state space towards the best possibilities for one goal and away from the best possibilities for another goal. This is *opportunity cost*.

Sometimes you gain more control over the future: most of the best possibilities make use of a windfall of cash. Sometimes you act to preserve control over the future: most Tic-Tac-Toe goals involve not ending the game right away. This is *power*.

Other people usually *objectively impact* you by decreasing or increasing a bunch of your AUs (generally, by changing your power). This happens for an extremely wide range of goals because of the structure of the environment.

Sometimes, the best possibilities are made unavailable or worsened only for goals very much like yours. This is *value impact*.

Sometimes a bunch of the best possibilities go through the same part of the future: fast travel to random places on Earth usually involves the airport. This is *instrumental convergence*.

*Exercise: Track what’s happening to your various AUs during the following story: you win the lottery. Being an effective spender, you use most of your cash to buy a majority stake in a major logging company. Two months later, the company goes under.*

## Technical appendix: AU landscape and world state contain equal information

In the context of finite deterministic Markov decision processes, there’s a wonderful handful of theorems which basically say that the AU landscape and the environmental dynamics encode each other. That is, they contain the *same* information, just with different emphasis. This supports thinking of the AU landscape as a “dual” of the world state.

Let be a rewardless deterministic MDP with finite state and action spaces , deterministic transition function , and discount factor . As our interest concerns optimal value functions, we consider only stationary, deterministic policies: .

The first key insight is to consider not policies, but the trajectories induced by policies from a given state; to not look at the state itself, but the

paths through timeavailable from the state. We concern ourselves with thepossibilitiesavailable at each juncture of the MDP.

To this end, for , consider the mapping of (where ); in other words, each policy maps to a function mapping each state to a discounted state visitation frequency vector , which we call a

possibility. The meaning of each frequency vector is: starting in state and following policy , what sequence of states do we visit in the future? States visited later in the sequence are discounted according to : the sequence would induce visitation frequency on , visitation frequency on , and visitation frequency on .

The possibility function outputs the possibilities available at a given state :

Put differently, the possibilities available are all of the potential film-strips of how-the-future-goes you can induce from the current state.

### Possibility isomorphism

We say two rewardless MDPs and are *isomorphic up to possibilities* if they induce the same possibilities. Possibility isomorphism captures the essential aspects of an MDP’s structure, while being invariant to state representation, state labelling, action labelling, and the addition of superfluous actions (actions whose results are duplicated by other actions available at that state). Formally, when there exists a bijection (letting be the corresponding -by- permutation matrix) satisfying for all .

This isomorphism is a natural contender^{[1]} for the canonical (finite) MDP isomorphism:

**Theorem:** and are isomorphic up to possibilities iff their directed graphs are isomorphic (and they have the same discount rate).

### Representation equivalence

Suppose I give you the following possibility sets, each containing the possibilities for a different state:

*Exercise: What can you figure out about the MDP structure? Hint: each entry in the column corresponds to the visitation frequency of a different state; the first entry is always , second , and third .*

You can figure out *everything*: , up to possibility isomorphism. Solution here.

How? Well, the norm of the possibility vector is always , so you can deduce easily. The single possibility state must be isolated, so we can mark that down in our graph. Also, it’s in the third entry.

The other two states correspond to the “1” entries in their possibilities, so we can mark that down. The rest follows straightforwardly.

**Theorem:** Suppose the rewardless MDP has possibility function . Given only ,^{[2]} can be reconstructed up to possibility isomorphism.

In MDPs, the “AU landscape” is the set of optimal value functions for all reward functions over states in that MDP. If you know the optimal value functions for just reward functions, you can also reconstruct the rewardless MDP structure.^{[3]}

From the environment (rewardless MDP), you can deduce the AU landscape (all optimal value functions) and all possibilities. From possibilities, you can deduce the environment and the AU landscape. From the AU landscape, you can deduce the environment (and thereby all possibilities).

*All of these encode the same mathematical object.*

## Technical appendix: Opportunity cost

Opportunity cost is when an action you take makes you more able to achieve one goal but less able to achieve another. Even this simple world has opportunity cost:

Going to the green state means you can’t get to the purple state as quickly.

On a deep level, why is the world structured such that this happens? Could you imagine a world without opportunity cost of any kind? The answer, again in the rewardless MDP setting, is simple: “yes, but the world would be trivial: you wouldn’t have any choices”. Using a straightforward formalization of opportunity cost, we have:

**Theorem:** Opportunity cost exists in an environment iff there is a state with more than one possibility.

Philosophically, opportunity cost exists when you have meaningful choices. When you make a choice, you’re necessarily moving away from some potential future but towards another; since you can’t be in more than one place at the same time, opportunity cost follows. Equivalently, we assumed the agent isn’t infinitely farsighted (); if it were, it would be possible to be in “more than one place at the same time”, in a sense (thanks to Rohin Shah for this interpretation).

While understanding opportunity cost may seem like a side-quest, each insight is another brick in the edifice of our understanding of the incentives of goal-directed agency.

### Notes

Just as game theory is a great abstraction for modelling competitive and cooperative dynamics, AU landscape is great for thinking about consequences: it automatically excludes irrelevant details about the world state. We can think about the effects of events without needing a specific utility function or ontology to evaluate them. In multi-agent systems, we can straightforwardly predict the impact the agents have on each other and the world.

“Objective impact to a location” means that agents whose plans route through the location tend to be objectively impacted.

The landscape is not the territory: AU is calculated with respect to an agent’s

*beliefs*, not necessarily with respect to what really “could” or will happen.

The possibility isomorphism is new to my work, as are all other results shared in this post. This apparent lack of basic theory regarding MDPs is strange; even stranger, this absence was actually pointed out in two published papers!

I find the existing MDP isomorphisms/equivalences to be pretty lacking. The details don’t fit in this margin, but perhaps in a paper at some point. If you want to coauthor this (mainly compiling results, finding a venue, and responding to reviews), let me know and I can share what I have so far (extending well beyond the theorems in my recent work on power). ↩︎

In fact, you can reconstruct the environment using only a limited subset of possibilities: the

*non-dominated*possibilities. ↩︎As a tensor, the transition function has size , while the AU landscape representation only has size . However, if you’re just representing as a transition

*function*, it has size . ↩︎

- Non-Obstruction: A Simple Concept Motivating Corrigibility by 21 Nov 2020 19:35 UTC; 63 points) (
- Attainable Utility Preservation: Empirical Results by 22 Feb 2020 0:38 UTC; 55 points) (
- Review of ‘But exactly how complex and fragile?’ by 6 Jan 2021 18:39 UTC; 51 points) (
- 2019 Review Rewrite: Seeking Power is Often Robustly Instrumental in MDPs by 23 Dec 2020 17:16 UTC; 35 points) (
- 6 Jan 2021 18:40 UTC; 33 points) 's comment on But exactly how complex and fragile? by (
- Reasons for Excitement about Impact of Impact Measure Research by 27 Feb 2020 21:42 UTC; 31 points) (
- Game-theoretic Alignment in terms of Attainable Utility by 8 Jun 2021 12:36 UTC; 16 points) (
- 23 Dec 2020 17:14 UTC; 12 points) 's comment on Seeking Power is Often Convergently Instrumental in MDPs by (
- 16 May 2020 17:32 UTC; 9 points) 's comment on Conclusion to ‘Reframing Impact’ by (
- 24 May 2020 20:55 UTC; 8 points) 's comment on Conclusion to ‘Reframing Impact’ by (
- 5 Feb 2021 0:09 UTC; 2 points) 's comment on Non-Obstruction: A Simple Concept Motivating Corrigibility by (
- 13 Apr 2020 15:35 UTC; 2 points) 's comment on “How conservative” should the partial maximisers be? by (

Values often seem back driven by beliefs about sensitivity of au. i.e. a person thinks that having 3x current money will be largest impact so they turn themselves into the sort of person who they think has the greatest chance of getting said money. This makes being wrong about sensitivity of outcomes to changes in various inputs scarier, since on expectation one should expect oneself to only do a middling job of keeping one’s identity disentangled from their strategies.

In a complete graph, all nodes are directly connected.

The opposite of this, is that if it were possible for an agent to be in more than one place at the same time, they could be infinitely farsighted. (Possibly as a consequence of FTL.)

Surprisingly, unless you’re talking about K1 (complete 1-graph), opportunity cost still exists in Kn (n>1). Each round, you choose where to go next (and you can go to any state immediately). Going to one state next round means you can’t go to a different state next round, so for any given action there exists a reward function which incurs opportunity cost.

Definition.We sayopportunity cost existsat a state s if there exist child states s1,s2 of state s such that V∗R(s1)≠V∗R(s2) for some reward function R. That is, s has successor states with different (optimal) AUs for some reward function.Things get weird here, depending on your theory of identity and how that factors into the planning / reward process? Can you spell this out some more?

There is a star, many light years away. If you exist in two locations at once simultaneously, from which the star is visible, and those two locations are not the same distance from the star, then intuitively, by seeing the star first from the closer position, you can know what it will look like from the second before it happens.

Less trivially, by altering the relative speeds of the two versions (with FTL telepathy), and setting up suitable devices for signaling, I think in theory, this would enable turning FTL into time travel. (Person A performs a calculation, and sends the results to Person B. Since Person A is the future version of person B, and they’re the same person in two places at once simultaneously, then by ‘de-synchronizing them right’ a message can be sent into the past.)

The technical appendix felt like it was more difficult than previous posts, but I had the advantage of having tried to read the paper from the preceding post yesterday and managed to reconstruct the graph & gamma correctly.

The early part is slightly confusing, though. I thought AU is a thing that belongs to the goal of an agent, but the picture made it look as if it’s part of the object (“how fertile is the soil?”). Is the idea here that the soil-AU is slang for “AU of goal ‘plant stuff here’”?

I did interpret the first exercise as “you planned to go onto the moon” and came up with stuff like “how valuable are the stones I can take home” and “how pleasant will it be to hang around.”

One thing I noticed is that the formal policies don’t allow for all possible “strategies.” In the graph we had to reconstruct, I can’t start at s1, then go to s1 once and then go to s3. So you could think of the larger set ΠL where the policies are allowed to depend on the time step. But I assume there’s no point unless the reward function also depends on the time step. (I don’t know anything about MDPs.)

Am I correct that a deterministic transition function is a function T:S×A→S and a non-deterministic one is a function T:S×A×S→[0,1]?

yes

yeah, this is because those are “nonstationary” policies—you change your mind about what to do at a given state. A classic result in MDP theory is that you never need these policies to find an optimal policy.

yup!

I have a paper on equivalences (and counterfactual equivalences, which is stronger) for POMDPs: https://arxiv.org/abs/1801.03737