Abductive reasoning is a well established field of Artificial Intelligence widely applied to different problem domains not least cognitive robotics and planning. It has been used to abduce high-level descriptions of the world from robot sense data, using rules that tell us what sense data would be generated by certain objects and events of the robots world, subject to certain constraints on their co-occurrence. It has also been used to abduce actions that might result in a desired goal state of the world, using descriptions of the normal effects of these actions, subject to constraints on the action combinations. We can generalise these applications to a multi-agent context. Several robots can collaboratively try to abduce an agreed higher-level description of the state of the world from their separate sense data consistent with their collective constraints on the abduced description. Similarly, multi-agent planning can be accomplished by the abduction of the actions of a collective plan where each agent uses its own description of the effect of its actions within the plan, such that the constraints on the actions of all the participating agents are satisfied.
To address this class of problems, we need to generalise the single agent abductive reasoning algorithm to a distributed abductive inference algortihm. In addition, if we want to investigate applications in which the set of collaborating robots/agents is open, we need an algorithm that allows agents to join or leave the collaborating group whilst a particular inference is under way, but which still produces sound abductive inferences. This paper describes such a distributed abductive reasoning system, which we call DARE, and its implementation in the multi-threaded Qu-Prolog variant of Prolog. We prove the soundness of the algorithm it uses and we discuss its completeness in relation to non-distributed abductive reasoning.
We illustrate the use of the algorithm with a multi-agent meeting scheduling example. The task is open in that the actual agents who need to attend is not determined in advance. Each individual agent has its own constraints on the possible meeting time and concerning which other agents must or must attend the meeting, if it attends. The algorithm selects the agents to attend and ensures that the constraints of each of the attending agents are satisfied.