Plan Assessment for Autonomous Manufacturing as Bayesian Inference ⋆ Paul Maier, Dominik Jain, Stefan Waldherr and Martin Sachenbacher Technische Universität München, Department of Informatics Boltzmanstraße 3, 85748 Garching, Germany {maierpa,jain,waldherr,sachenba}@in.tum.de Abstract. Next-generation autonomous manufacturing plants create individual- ized products by automatically deriving manufacturing schedules from design specifications. However, because planning and scheduling are computationally hard, they must typically be done offline using a simplified system model, mean- ing that online observations and potential component faults cannot be considered. This leads to the problem of plan assessment: Given behavior models and current observations of the plant’s (possibly faulty) behavior, what is the probability of a partially executed manufacturing plan succeeding? In this work, we propose 1) a statistical relational behavior model for a class of manufacturing scenarios and 2) a method to derive statistical bounds on plan success probabilities for each prod- uct from confidence intervals based on sampled system behaviors. Experimental results are presented for three hypothetical yet realistic manufacturing scenarios. 1 Introduction In a scenario of mass customization using autonomous manufacturing, a factory is en- visaged that generates, during the night, the manufacturing plans for numerous indi- vidualized products to be produced the next day. It employs model-based planning and scheduling capabilities, which use very abstract models to keep planning/scheduling tractable, omitting e.g. behavioral knowledge about potential failures of factory stations. In addition, observations made at execution time are not available at planning/scheduling time. In the light of such partial observations, it may become clear that certain plans will fail, e.g. if a plan operates a component that is now likely to be faulty. This leads to a problem of evaluating manufacturing plans with respect to online observations, based on models focussed on station behavior. It is especially interesting from the point of view of autonomous manufacturing control, where systems are rigid enough to allow automated advance planning/scheduling (rather than online planning), yet bear inherent uncertainties such as station failures. We call this evaluation plan assessment [1]. The idea is to compute, for each prod- uct, bounds on the respective success probability. This allows to decide whether to 1) continue with a plan, 2) stop the plan because it probably will not succeed or 3) gather more information. It requires a) models of the complex, uncertain interactions among products and factory stations and b) efficient reasoning. In [1] we proposed using probabilistic automata models and a solution based on constraint optimization, ⋆ Preprint submitted to KI 2010.