Earlier we established that the quantilal policy can be computed in polynomial time to any given approximation (see "Proposition 5"). Now we show that an exact quantilal policy can be computed in polynomial time (in particular there is always a rational quantilal policy).

We assume geometric time discount throughout.

Lemma

Consider . Define the linear operators and by

(Note that this is the transpose of the defined in "Proposition A.3" of the previous essay.)

Then, if and only if there is s.t.

Proof

(This is actually well known, but we spell out the proof to be self-contained.)

Suppose that . We already know that this implies that there is a stationary policy s.t. (we abuse notation in the obvious way): see the proofs of "Proposition 2" and "Proposition 3". Define the linear operator by

It follows that

Define by

We have

Also, obviously . We get

Conversely, suppose that is as above. Since , there is s.t. for any , if then

Again, we have

Also, for the same reason as before

By the assumption, the left hand side equals . We conclude

Theorem

Assuming all parameters are rational like before, there is a polynomial time algorithm that computes a quantilal policy.

Proof

The algorithm starts by solving the following linear program. The indeterminates are and . The goal is maximizing . The constraints are

Then, the algorithm computes s.t. for any , if then

For s.t. , is arbitrary.

Now we need to explain why this algorithm is correct.

Observe that, the first 3 constraints mean that defined by lies in (by Lemma 1) and, moreover, for s.t. . It remains to show that the linear program amounts to maximizing inside . Indeed, the 4th constraint just means that . The last constraint implies that we are actually maximizing

The latter is indeed , since every corresponds to a pure strategy of the adversary in the corresponding zero-sum game: namely, this strategy is setting the penalty function to

(Strategies that place non-vanishing penalty on states outside of become irrelevant after imposing the 4th constraint. The remaining penalty functions form a simplex with vertices as above.)

QuantilizationAI
Personal Blog
New Comment