Intelligent Agent Foundations Forumsign up / log in
Logical Induction with incomputable sequences
post by Alex Mennen 185 days ago | Abram Demski, Paul Christiano and Scott Garrabrant like this | discuss

In the definition of a logical inductor, the deductive process is required to be computable. This, of course, does not allow the logical inductor to use randomness, or predict uncomputable sequences. The way traders were defined in the logical induction paper, this was necessary, because the traders were not given access to the output of the deductive process.

To fix this, a trading strategy for day \(n\) should be redefined as a function that takes in the output of the deductive process on day \(n-1\) as its input, and outputs what the logical induction paper defines as a trading strategy for day \(n\); that is, an affine combination of the form \(c+\xi_{1}\phi_{1}+...+\xi_{k}\phi_{k}\), where \(\phi_{1},...,\phi_{k}\) are sentences, \(\xi_{1},...,\xi_{k}\) are expressible features of rank \(\leq n\), and \(c=-\sum_{i}\xi_{i}\phi_{i}^{*n}\). A trader is a function which takes in \(n\) and outputs a trading strategy for day \(n\). By Currying, a trader can be seen as a function that takes in a number \(n\) and a list of sentences given by the deductive process, and outputs an expressible feature combination as above. We can say that a trader is efficiently computable if this function is computable in time polynomial in \(n\) plus the total length of the sentences output by the deductive process. The definition of exploitation would be modified in the natural way, and there is also a natural way to modify the logical induction algorithm, which will satisfy the logical induction criterion.

As an example, suppose a logical inductor is given access to a sensor that regularly produces bits based on what it observes in the environment. We can represent the data from the sensor with an additional unary predicate \(S\) that we add to the language, such that \(S\left(n\right)\) is true iff the \(n\)th bit provided by the sensor is a \(1\) (this assumes that we’re working in a theory that can interpret arithmetic, so that ``\(n\)’’ can be expressed in the language). The deductive process should output \(S\left(n\right)\) or \(\neg S\left(n\right)\) on day \(n\) (and also can output consequences that it can deduce from the values of the bits it has seen so far). Or, if the logical inductor gets access to more empirical information or random bits as time goes on, there could be an increasing function \(f\) such that the deductive process outputs the truth values of \(S\left(f\left(n\right)\right),...,S\left(f\left(n+1\right)-1\right)\) on day \(n\). Note that in this situation, the deductive process is computable as a function of the bitstream given by the sensor, so the traders may as well take in as input only the bits from the sensor that the deductive process has seen by day \(n-1\), rather than every sentence produced by the deductive process.

This seems to be similar to what Vadim was doing in section 3 of this paper, except that that paper moved to a continuous setting, did not deal with computability, and abandoned predicting theorems as a goal.



NEW LINKS

NEW POSTS

NEW DISCUSSION POSTS

RECENT COMMENTS

[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

RSS

Privacy & Terms