opening it up with Common Lisp
Book review: Darwinia
Summer reading: Spin
the Omnivoire's Delimma
the Golem's Eye
A New Framework for Sensor Interpretation
Titles like this one remind me why you should never call something new or complete or best: time moves on and it starts to sound strange!
Carver and Lesser present one of the final words on control from of the Blackboard heyday of the late 1980s and the early 1990s. The Blackboard architecture is a framework for thinking about problem solving as an opportunistic, cooperative, and flexible process. It was born on work in uncertain, hypothesis rich domains like speech understanding and signal processing. It faded for a variety of mostly sociological reasons since, truth told, there really isn't a better architecture out there. Like Lisp, Blackboards are seeing something of a resurgence of late (there is an even an open source version of the de-facto Blackboard standard being worked on) -- there is yet hope in the world.
Metaphorically speaking, a Blackboard is a shared memory space where lots of experts do their brainstorming. Each expert works on part of the problem and all contribute when and where they can. The trouble is that if all the experts try to work at once and present all their ideas, the Blackboard will become a mess and nothing useful will get done. The control problem is that of figuring which expert(s) ought to have their say next.
The RESUN framework presented in this paper says that control should be based on "gathering evidence to resolve particular sources of uncertainty." RESUN implements this via a script based planner that can return (refocus) to previous decisions (and re-decide) as new evidence comes in. These plans provide the context (goal/plan/sub-goal) for the activities of the Blackboard. The nice thing about RESUN is that the Blackboard can do more than just make hypotheses and test them, it can also do differential diagnosis (if A and B are competing hypotheses for my data, then I can get evidence for A by finding evidence against B).
The work presented here still feels quite fresh even after 13-years. The Blackboard control problem was never really solved -- doing so is probably AI complete -- but solutions like this one show both that there is unexplored potential and that much work remains.
Copyright -- Gary Warren King, 2004 - 2006