Logical thermodynamics: towards a theory of self-trusting uncertain reasoning

2Manfred

2Squark

0Manfred

2Luke_A_Somers

New Comment

4 comments, sorted by Click to highlight new comments since: Today at 6:36 PM

A general problem with big integrals in your definition of the abstract function P: Suppose that C is a statement whose shortest proof takes longer to write out than your agent has time. When asked for P(C), your abstract function P should not assign a number that your agent is physically incapable of assigning.

More concrete problem: since your energies are all greater than 0, proofs (evidence) with long length get weighted greater (are lower energy) than short proofs.

Evidence with long length is weighted *less* since it contributes less to the energy.

When computing probabilities at finite temperature to finite precision, it should be possible to ignore long evidence (at least evidence which is long wrt the given sentence). If the shortest evidence for a sentence is very long, it means its probability is "marginal" in some sense (since the total energy depends weakly on the value of the probability field at this statement). For sentences which cannot be decomposed into simpler sentences using propositional calculus operations, "marginal" probably means close to 1/2.

Followup to: Overcoming the Loebian obstacle using evidence logic

In the previous post I proposed a probabilistic system of reasoning for overcoming the Loebian obstacle. For a consistent theory it seems natural the expect such a system should yield a coherent probability assignment in the sense of Christiano et al. This means that

a. provably true sentences are assigned probability 1

b. provably false sentences are assigned probability 0

c. The following identity holds for any two sentences φ, ψ

[1] P(φ) = P(φ and ψ) + P(φ and not-ψ)

In the previous formalism, conditions a & b hold but condition c is violated (at least I don't see any reason it should hold).

In this post I attempt to achieve the following:

## Construction

Define a

probability fieldto be a function p : {sentences} -> [0, 1] satisfying the following conditions:in propositional calculus(e.g. φ = ψ or not-ψ) then p(φ) = 1energyof a probability field p to be E(p) := Σ_{φ}Σ_{v}2^{-l(v) }E_{φ,v}(p(φ)). Herevare pieces of evidence as defined in the previous post, E_{φ,v}are their associated energy functions and l(v) is the length of (the encoding of)v. We assume that the encoding ofvcontains the encoding of the sentence φ for which it is evidence and E_{φ,v}(p(φ)) := 0 for all φ except the relevant one. Note that the associated energy functions are constructed in the same way as in the previous post, however they arenotthe same because of the self-referential nature of the construction: it refers to final probability assignment.The final probability assignment is defined to be

P(φ) = Integral

_{p}[e^{-E(p)/T }p(φ)] / Integral_{p}e^{-E(p)/T}Here T >= 0 is a parameter representing the magnitude of logical uncertainty. The integral is infinite-dimensional so it's not obviously well-defined. However, I suspect it can be defined by truncating to a finite set of statements and taking a limit wrt this set. In the limit T -> 0, the expression should correspond to computing the centroid of the set of minima of E (which is convex because E is convex).

## Remarks