Abductive reasoning is a powerful logic inference mechanism that allows assumptions to be
made during answer computation for a query, and thus is suitable for reasoning over incomplete
knowledge. Multi-agent hypothetical reasoning is the application of abduction in a distributed
setting, where each computational agent has its local knowledge representing partial world and
the union of all agents' knowledge is still incomplete. It is different from simple distributed
query processing because the assumptions made by the agents must also be consistent with
global constraints.
Multi-agent hypothetical reasoning has many potential applications, such as collaborative planning
and scheduling, distributed diagnosis and cognitive perception. Many of these applications
require the representation of arithmetic constraints in their problem specifications as well as
constraint satisfaction support during the computation. In addition, some applications may
have confidentiality concerns as restrictions on the information that can be exchanged between
the agents during their collaboration. Although a limited number of distributed abductive systems
have been developed, none of them is generic enough to support the above requirements.
In this thesis we develop, in the spirit of Logic Programming, a generic and extensible distributed
abductive system that has the potential to target a wide range of distributed problem
solving applications. The underlying distributed inference algorithm incorporates constraint
satisfaction and allows non-ground conditional answers to be computed. Its soundness and
completeness have been proved. The algorithm is customisable in that different inference and
coordination strategies (such as goal selection and agent selection strategies) can be adopted
while maintaining correctness. A customisation that supports confidentiality during problem
solving has been developed, and is used in application domains such as distributed security
policy analysis. Finally, for evaluation purposes, a
flexible experimental environment has been
built for automatically generating different classes of distributed abductive constraint logic programs.
This environment has been used to conduct empirical investigation of the performance
of the customised system.