A tractable, interpretable formulation of approximate conditioning for pairwise-specified probability distributions over truth values discussion post by Janos Kramar 992 days ago | Victoria Krakovna and Patrick LaVictoire like this | 2 comments These results from my conversations with Charlie Steiner at the May 29-31 MIRI Workshop on Logical Uncertainty will primarily be of interest to people who’ve read section 2.4 of Paul Christiano’s Non-Omniscience paper. If we write a reasoner that keeps track of probabilities of a collection of sentences $$\varphi_1,\dots,\varphi_n$$ (that grows and shrinks as the reasoner explores), we need some way of tracking known relationships between the sentences. One way of doing this is to store the pairwise probability distributions, ie not only $$\Pr(\varphi_i)$$ for all $$i$$ but also $$\Pr(\varphi_i\wedge\varphi_j)$$ for all $$i,j$$. If we do this, a natural question to ask is: how can we update this data structure if we learn that eg $$\varphi_1$$ is true? We’ll refer to the updated probabilities as $$\Pr(\cdot|\varphi_1)$$. It’s fairly reasonable for us to want to set $$\Pr(\varphi_i|\varphi_1):=\Pr(\varphi_i\wedge\varphi_1)/\Pr(\varphi_1)$$; however, it’s less clear what values to assign to $$\Pr(\varphi_i\wedge\varphi_j|\varphi_1)$$, because we haven’t stored $$\Pr(\varphi_i\wedge\varphi_j\wedge\varphi_1)$$. One option would be to find the maximum entropy distribution over truth assignments to $$\varphi_1,\dots,\varphi_n$$ under the constraint that the stored pairwise distributions are correct. This seems intractable for large $$n$$; however, in the spirit of locality, we could restrict our attention to the joint truth value distribution of $$\varphi_1,\varphi_i,\varphi_j$$. Maximizing its entropy is simple (it boils down to either convex optimization or solving a cubic), and yields a plausible candidate for $$\Pr(\varphi_i\wedge\varphi_j\wedge\varphi_1)$$ that we can derive $$\Pr(\varphi_i\wedge\varphi_j|\varphi_1)$$ from. I’m not sure what global properties this has, for example whether it yields a positive semidefinite matrix $$(\Pr(\varphi_i\wedge\varphi_j))_{i,j}$$. A different option, as noted in section 2.4.2, is to observe that the matrix $$(\Pr(\varphi_i\wedge\varphi_j))_{i,j}$$ must be positive semidefinite under any joint distribution for the truth values. This means we can consider a zero-mean multivariate normal distribution with this matrix as its covariance; then there’s a closed-form expression for the Kullback-Leibler divergence of two such distributions, and this can be used to define a sort of conditional distribution, as is done in section 2.4.3. However, as the paper remarks, this isn’t a very familiar way of defining these updated probabilities. For example, it lacks the desirable property that $$\Pr(\varphi_i|\varphi_1)=\Pr(\varphi_i\wedge\varphi_1)/\Pr(\varphi_1)$$. Fortunately, there is a natural construction that combines these ideas: namely, if we consider the maximum-entropy distribution for the truth assignment vector $$(1_{\varphi_1},\dots,1_{\varphi_n})$$ with the given second moments $$\operatorname{E}(1_{\varphi_i}1_{\varphi_j})$$, but relax the requirement that their values be in $$\{0,1\}$$, then we find a multivariate normal distribution $\mathcal{N}\left(\left(\Pr(\varphi_i)\right)_i,\left(\Pr(\varphi_i\wedge\varphi_j)-\Pr(\varphi_i)\Pr(\varphi_j)\right)_{i,j}\right)\mbox{.}$ If we wish to update this distribution after observing $$\varphi_1$$ by finding the candidate distribution $$(1_{\varphi_1},\dots,1_{\varphi_n}|\varphi_1)$$ of highest relative entropy with $$\Pr(1_{\varphi_1}=1|\varphi_1)=1$$, as proposed in the paper, then we will get the multivariate normal conditional distribution $\mathcal{N}\left(\left(\Pr(\varphi_1\wedge\varphi_i)/\Pr(\varphi_1)\right)_i,\left(\Pr(\varphi_i\wedge\varphi_j)-\Pr(\varphi_i)\Pr(\varphi_j)-\frac{(\Pr(\varphi_1\wedge\varphi_i)-\Pr(\varphi_1)\Pr(\varphi_i))(\Pr(\varphi_1\wedge\varphi_j)-\Pr(\varphi_1)\Pr(\varphi_j))}{\Pr(\varphi_1)-\Pr(\varphi_1)^2}\right)_{ij}\right)\mbox{.}$ Note that this generally has $$\operatorname{Var}\left(1_{\varphi_i}\middle|\varphi_1\right)\not=\operatorname{E}\left(1_{\varphi_i}\middle|\varphi_1\right)\left(1-\operatorname{E}\left(1_{\varphi_i}\middle|\varphi_1\right)\right)$$, which is a mismatch; this is related to the fact that a conditional variance in a multivariate normal is never higher than the marginal variance, which is an undesirable feature for a distribution over truth-values. This is also related to other undesirable features; for example, if we condition on more than one sentence, we can arrive at conditional probabilities outside of $$[0,1]$$. (For example if 3 sentences have $$\Pr(\varphi_1)=\Pr(\varphi_2)=\Pr(\varphi_3)=\frac{1}{3},\Pr(\varphi_1\wedge\varphi_2)=\Pr(\varphi_1\wedge\varphi_3)=\Pr(\varphi_2\wedge\varphi_3)=\varepsilon$$ then this yields $$\Pr(\varphi_3|\varphi_1,\varphi_2)=\frac{-1+15\varepsilon}{1+9\varepsilon}\approx -1$$; this makes sense because this prior is very confident that $$1_{\varphi_1}+1_{\varphi_2}+1_{\varphi_3}\approx 1$$, with standard deviation $$\sqrt{6\varepsilon}$$.) Intermediate relaxations that lack these particular shortcomings are possible, such as the ones that restrict the relaxed $$1_{\varphi_1},\dots,1_{\varphi_n}$$ to the sphere $$\sum_i (2x_i-1)^2=n$$ or ball $$\sum_i (2x_i-1)^2\leq n$$. Then the maximum entropy distribution, similarly to a multivariate normal distribution, has quadratic logdensity, though the Hessian of the quadratic may have nonnegative eigenvalues (unlike in the normal case). In the spherical case, this is known as a Fisher-Bingham distribution. Both of these relaxations seem difficult to work with, eg to compute normalizing constants for; furthermore I don’t think the analogous updating process will share the desirable property that $$\Pr(\varphi_i|\varphi_1)=\Pr(\varphi_i\wedge\varphi_1)/\Pr(\varphi_1)$$. However, the fact that these distributions allow updating by relaxed conditioning, keep (fully conditioned) truth-values between 0 and 1, and have reasonable (at least, possibly-increasing) behavior for conditional variances, makes them seem potentially appealing.

 by Janos Kramar 983 days ago | Jessica Taylor likes this | link An easy way to get rid of the probabilities-outside-[0,1] problem in the continuous relaxation is to constrain the “conditional”/updated distribution to have $$\operatorname{Var}\left(1_{\varphi_i}\middle|\dots\right)\leq \operatorname{E}\left(1_{\varphi_i}\middle|\dots\right)\left(1-\operatorname{E}\left(1_{\varphi_i}\middle|\dots\right)\right)$$ (which is a convex constraint; it’s equivalent to $$\operatorname{Var}\left(1_{\varphi_i}\middle|\dots\right)+\left(\operatorname{E}\left(1_{\varphi_i}\middle|\dots\right)-\frac{1}{2}\right)^2\leq \frac{1}{4}$$), and then minimize KL-divergence accordingly. The two obvious flaws are that the result of updating becomes ordering-dependent (though this may not be a problem in practice), and that the updated distribution will sometimes have $$\operatorname{Var}\left(1_{\varphi_i}\middle|\dots\right)< \operatorname{E}\left(1_{\varphi_i}\middle|\dots\right)\left(1-\operatorname{E}\left(1_{\varphi_i}\middle|\dots\right)\right)$$, and it’s not clear how to interpret that. reply
 by Janos Kramar 974 days ago | Patrick LaVictoire likes this | link There is a lot more to say about the perspective that isn’t relaxed to continuous random variables. In particular, the problem of finding the maximum entropy joint distribution that agrees with particular pairwise distributions is closely related to Markov Random Fields and the Ising model. (The relaxation to continuous random variables is a Gaussian Markov Random Field.) It is easily seen that this maximum entropy joint distribution must have the form \(\log \Pr(1_{\varphi_1},\dots,1_{\varphi_n})=\sum_{i

NEW DISCUSSION POSTS

[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