Planning for modular printers: beyond productivity

Abstract

This paper reports our experience extending an on-line printer controller based on AI planning to handle two significant features of our real-world domain: execution failures and multi-objective preferences. A printer controller must plan quickly, otherwise expensive human intervention will be required. Our approach is practical and efficient, and showcases the flexibility inherent in viewing planning as heuristic search. Execution failure is handled by replanning. We link together the individual searches for each in-flight sheet, giving rise to a tree of potentially infinite branching factor. Multiple objectives are handled by using multiple pre-computed pattern databases to compute scores that control tie-breaking during best-first search. Our experiments on multiple prototype printing systems show that replanning and preference-handling can be made practical without using hand-coded control knowledge.