DOI:
Abstract:
DTGolog, a decision-theoretic agent programming language based on the situation calculus, was proposed to ease some of the computational difficulties associated with Markov Decision Processes (MDPs) by using natural ordering constraints on execution of actions. Using DTGolog, domain specific constraints on a set of policies can be expressed in a high-level program to reduce significantly computations required to find a policy optimal in this set. We explore whether the DTGolog framework can be used to evaluate different designs of a decision making agent in a large real-world domain. Each design is understood as combination of a template (expressed as a Golog program) for available policies and a reward function. To evaluate and compare alternative designs we estimate the probability of goal satisfaction for each design. As a domain, we choose the London Ambulance Service (LAS) case study that is well known in software engineering, but remains unknown in AI. We demonstrate that DTGolog can be applied successfully to quantitative evaluation of alternative designs in terms of their ability to satisfy a system goal with a high probability. The full version of this paper includes a detailed axiomatization of the domain in the temporal situation calculus with stochastic actions. The main advantage of this representation is that neither actions, nor states require explicit enumeration. We do an experimental analysis using an on-line implementation of DTGolog coupled with a simulator that models real time actions of many external agents.