Intelligent Agent Foundations Forumsign up / log in
Optimal and Causal Counterfactual Worlds
post by Scott Garrabrant 1018 days ago | Sam Eisenstat, Abram Demski, Daniel Dewey, Nate Soares and Patrick LaVictoire like this | 2 comments

Let \(L\) denote the language of Peano arithmetic. A (counterfactual) world \(W\) is any subset of \(L\). These worlds need not be consistent. Let \(\mathcal{W}\) denote the set of all worlds. The actual world \(W_\mathbb{N}\in\mathcal{W}\) is the world consisting of all sentences that are true about \(\mathbb{N}\).

Consider the function \(C:L\rightarrow\mathcal{W}\) which sends the sentence \(\phi\) to the world we get by “correctly” counterfactually assuming \(\phi\). The function \(C\) is not formally defined, because we do not yet have a satisfactory theory of logical counterfactuals.

Hopefully we all agree that \(\phi\in C(\phi)\) and \(\phi\in W_\mathbb{N}\Rightarrow C(\phi)=W_\mathbb{N}\).

Given an (infinite) directed acyclic graph \(G\), and a map \(v\) from sentences to vertices of \(G\), we say that \(C\) is consistent with \(G\) and \(v\) if \(C(\phi)=C(\psi)\) for all \(v(\phi)=v(\psi)\), and whenever \(W_{\mathbb{N}}\) and \(C(\phi)\) disagree on a sentence \(\psi\) there must exist some causal chain \(\psi_1,\ldots \psi_n\) such that:

  1. \(v(\psi_1)=v(\phi)\),

  2. \(\psi_n=\psi\),

  3. \(W_{\mathbb{N}}\) and \(C(\phi)\) disagree on every \(\psi_i\), and

  4. \(v(\psi_i)\) is a parent of \(v(\psi_{i+1})\).

These conditions give a kind of causal structure such that changes from \(W_{\mathbb{N}}\) and \(C(\phi)\) must propagate through the graph \(G\).

Given a function \(f:\mathcal{W}\rightarrow\mathbb{R}\), we say that \(C\) optimizes \(f\) if for all \(\phi\in W\) and \(W\neq C(\phi)\) we have \(f(C(\phi))>f(W)\).

Many approaches to logical counterfactuals can be described either as choosing the optimal world (under some function) in which \(\phi\) is true or observing the causal consequences of setting \(\phi\) to be true. The purpose of this post is to prove that these frameworks are actually equivalent, and to provide a strategy for possibly showing that no attempt at logical counterfactuals which could be described within either framework could ever be what we mean by “correct” logical counterfactuals.

A nontrivial cycle in \(C\) is a list of sentences \(\phi_1,\ldots,\phi_n\), such that \(\phi_{i}\in C(\phi_{i+1})\), \(\phi_n\in C(\phi_1)\), and the worlds \(C(\phi_i)\) are not all the same for all \(i\).

Given a partial order \(\succ\) , we say that \(C\) optimizes \(\succ\) if for all \(\phi\in W\) and \(W\neq C(\phi)\) we have \(C(\phi)\succ W\).

Our main result is that the following are equivalent:

  1. \(C\) optimizes \(f\) for some function \(f\).

  2. \(C\) is consistent with \(G\) and \(v\) for some DAG \(G\) and map \(v\).

  3. \(C\) has no nontrivial cycles.

  4. \(C\) optimizes \(\succ\) for some partial order \(\succ\).


1 \(\Rightarrow\) 2: Construct the graph \(G\) with a vertex for every world in the image of \(C\). The map \(v\) sends \(\phi\) to the vertex associated with \(C(\phi)\). Insert an edge to \(W_{\mathbb{N}}\) from every other vertex. Insert an edge from the vertex associated with \(W_1\neq W_{\mathbb{N}}\) to the vertex associated with \(W_2\neq W_{\mathbb{N}}\) whenever \(f(W_1)<f(W_2)\). Clearly \(C(\phi)=C(\psi)\) for all \(v(\phi)=v(\psi)\).

Assume \(W_{\mathbb{N}}\) and \(C(\phi)\) disagree on a sentence \(\psi\). If \(\psi\in W_{\mathbb{N}}\) then \(v(\phi)\rightarrow v(\psi)\), since \(v(\psi)\) is the vertex associated with \(W_\mathbb{N}\) which is a child of every vertex. Therefore you get a length 1 path from \(v(\phi)\) to \(v(\psi)\). Otherwise, \(\psi\notin W_{\mathbb{N}}\), so \(\psi\in C(\phi)\). In this case, note that since \(\psi\) is in both \(C(\phi)\) and \(C(\psi)\), and these worlds are not the same, it must be that \(f(C(\psi))>f(C(\phi))\). Again, this means that you get a length 1 path from \(v(\phi)\) to \(v(\psi)\). Therefore \(C\) is consistent with \(G\) and \(v\).

2 \(\Rightarrow\) 3: Consider a nontrivial cycle \(\phi_1,\ldots,\phi_n\). If any of these sentences were in \(W_\mathbb{N}\), then they would all be true in \(W_\mathbb{N}\), since \(C(\phi)=W_\mathbb{N}\) whenever \(\phi\in W_\mathbb{N}\). This would contradict the fact that \(\phi_1,\ldots,\phi_n\) is a nontrivial cycle.

Otherwise, since \(\phi_i\in C(\phi_{i+1})\), there must be a path from \(v(\phi_{i+1})\) to \(v(\phi_i)\) in \(G\), since \(C(\phi_{i+1})\) and \(W_\mathbb{N}\) differ on \(\phi_i\). Concatenating these paths together would give a cycle in \(G\) unless the \(v(\phi_i)\) are all the same vertex. However, if all of the \(v(\phi_i)\) are the same vertex, then all of the \(C(\phi_i)\) would be the same world, which would contradict the fact that \(\phi_1,\ldots,\phi_n\) is a nontrivial cycle.

3 \(\Rightarrow\) 4: Consider the partial order on the image of \(C\) constructed by saying that \(W_1\succ W_2\) if \(W_1=C(\phi)\) for some \(\phi\in W_2\), and taking the transitive closure of these rules. If this were not a partial order, it would have to be because we created a cycle of worlds \(W_1,\ldots,W_n\), such that each \(W_i=C(\phi_i)\) with \(\phi_i\in W_{i+1}\) and \(\phi_n\in W_{1}.\) This is would be a nontrivial cycle.

Extend this partial order to all of \(\mathcal{W}\) by saying that if \(W_1\) is in the image of \(C\) and \(W_2\) is not, then \(W_1\succ W_2\). Note that this \(C\) optimizes this partial order by definition.

4 \(\Rightarrow\) 1: Let \(C\) optimize the partial order \(\succ\). Consider the restriction of \(\succ\) to the image of \(C\). This is a partial order on a countable set. Order the worlds in this partial order \(W_1,W_2,\ldots.\) Embed the partial order into \(\mathbb{R}\) by repeatedly defining \(f(W_n)\) such that:

  1. \(f(W_n)>0\),

  2. \(f(W_n)>f(W_i)\) for all \(i<n\) and \(W_n\succ W_i\), and

  3. \(f(W_n)<f(W_i)\) for all \(i<n\) and \(W_i\succ W_n\).

Define \(f(W)=0\) if \(W\) is not in the image of \(C\). Clearly this function \(f\) is constructed such that C optimizes \(f\).


This result is useful not just for establishing the equivalence of (a certain type of) optimization and acyclic causal networks, but also for providing a strategy for showing that “correct” logical counterfactuals cannot arise as an optimization process or through an acyclic causal network. To show this, one only has to exhibit a single nontrivial cycle. In the simplest case, this can be done by exhibiting a pair of sentences \(\phi\) and \(\psi\) such that \(\phi\) and \(\psi\) are clearly counterfactual consequences of each other, but do not correspond to identical counterfactual worlds.

Do the “correct” logical counterfactuals exhibit a nontrivial cycle?

by Charlie Steiner 1016 days ago | Nate Soares and Patrick LaVictoire like this | link

This is an interesting way of thinking about logical counterfactuals. It all seems to come down to what desiderata you desiderate.

We might assign a DAG to \(W_\mathbb{N}\) by choosing a reference theorem-prover, which uses theorems/syntactic rules to generate more theorems. We then draw an edge to each sentence in \(W_\mathbb{N}\) from its direct antecedents in it first proof in the reference theorem-prover. One option is that \(C(\phi)\) would only be allowed to disagree with sentences in \(W_\mathbb{N}\) for sentences that are descendants of \(\neg \phi\). But this doesn’t specify your \(G\), because it doesn’t assign a place in the graph to sentences not in \(W_\mathbb{N}\).

If we try a similar theorem-prover assignment to specify \(C(\phi)\) with \(\phi\) false under PA, we’ll get very silly things as soon as the theorem-prover proves \(\perp\) and explodes; our graph will no longer follow the route analogous to the one for \(W_\mathbb{N}\). Is there some way to enforce that analogy?

Ideally what I think I desiderate is more complicated than only disagreeing on \(W_\mathbb{N}\)-descendants of \(\neg \phi\) - for example if \(\phi\) is a counterexample to a universal statement that is not a descendant of any individual cases, I’d like the universal statement to be counterfactually disproved.


by Sam Eisenstat 947 days ago | link

Condition 4 in your theorem coincides with Lewis’ account of counterfactuals. Pearl cites Lewis, but he also criticizes him on the ground that the ordering on worlds is too arbitrary. In the language of this post, he is saying that condition 2 arises naturally from the structure of the problem and that condition 4 is derives from the deeper structure corresponding to condition 2.

I also noticed that the function \(f\) and the partial order \(\succ\) can be read as “time of first divergence from the real world” and “first diverges before”, respectively. This makes the theorem a lot more intuitive.






[Delegative Reinforcement
by Vadim Kosoy on Stable Pointers to Value II: Environmental Goals | 1 like

Intermediate update: The
by Alex Appel on Further Progress on a Bayesian Version of Logical ... | 0 likes

Since Briggs [1] shows that
by 258 on In memoryless Cartesian environments, every UDT po... | 2 likes

This doesn't quite work. The
by Nisan Stiennon on Logical counterfactuals and differential privacy | 0 likes

I at first didn't understand
by Sam Eisenstat on An Untrollable Mathematician | 1 like

This is somewhat related to
by Vadim Kosoy on The set of Logical Inductors is not Convex | 0 likes

This uses logical inductors
by Abram Demski on The set of Logical Inductors is not Convex | 0 likes

Nice writeup. Is one-boxing
by Tom Everitt on Smoking Lesion Steelman II | 0 likes

Hi Alex! The definition of
by Vadim Kosoy on Delegative Inverse Reinforcement Learning | 0 likes

A summary that might be
by Alex Appel on Delegative Inverse Reinforcement Learning | 1 like

I don't believe that
by Alex Appel on Delegative Inverse Reinforcement Learning | 0 likes

This is exactly the sort of
by Stuart Armstrong on Being legible to other agents by committing to usi... | 0 likes

When considering an embedder
by Jack Gallagher on Where does ADT Go Wrong? | 0 likes

The differences between this
by Abram Demski on Policy Selection Solves Most Problems | 1 like

Looking "at the very
by Abram Demski on Policy Selection Solves Most Problems | 0 likes


Privacy & Terms