Published online by Cambridge University Press: 04 March 2009
Dynamic programming is a strategy for solving optimisation problems. In this paper, we show how many problems that may be solved by dynamic programming are instances of the same abstract specification. This specification is phrased using the calculus of relations offered by topos theory. The main theorem underlying dynamic programming can then be proved by straightforward equational reasoning.
The generic specification of dynamic programming makes use of higher-order operators on relations, akin to the fold operators found in functional programming languages. In the present context, a data type is modelled as an initial F-algebra, where F is an endofunctor on the topos under consideration. The mediating arrows from this initial F-algebra to other F-algebras are instances of fold – but only for total functions. For a regular category ε, it is possible to construct a category of relations Rel(ε). When a functor between regular categories is a so-called relator, it can be extended (in some canonical way) to a functor between the corresponding categories of relations. Applied to an endofunctor on a topos, this process of extending functors preserves initial algebras, and hence fold can be generalised from functions to relations.
It is well-known that the use of dynamic programming is governed by the principle of optimality. Roughly, the principle of optimality says that an optimal solution is composed of optimal solutions to subproblems. In a first attempt, we formalise the principle of optimality as a distributivity condition. This distributivity condition is elegant, but difficult to check in practice. The difficulty arises because we consider minimum elements with respect to a preorder, and therefore minimum elements are not unique. Assuming that we are working in a Boolean topos, it can be proved that monotonicity implies distributivity, and this monotonicity condition is easy to verify in practice.