DÉJÀ VU – A Reusable Framework for the Construction of Intelligent Interactive Schedulers Jürgen Dorn, Mario Girsch, and Nikos Vidakis Vienna University of Technology Christian Doppler Laboratory for Expert Systems Paniglgasse 16, A-1040 Vienna, Austria E-mail {dorn|girsch|vidakis}@dbai.tuwien.ac.at ABSTRACT We describe the basic techniques underlying the DÉJÀ VU Scheduling Class Library to achieve a library of reusable and extendible classes for the construction of interactive production scheduling systems. The constructed systems shall be efficient and user centered which means that the user shall have full control over the schedule construction process. Mixed-initiative scheduling shall be possible. We present how scheduling objects and constraints on these objects are realized. Further we describe the potential user interactions with the system and show prototypical examples from the graphical user interface. A first scheduling system was developed with this class library for the steel plant of Böhler in Kapfenberg. We demonstrate which extensions we made for this system. 1. DESIGN PRINCIPLES OF DÉJÀ VU DÉJÀ VU is a framework of C++ classes to support devel- opers in the construction of scheduling systems for indus- trial production processes. The design of the framework was directed by the following criteria: the scheduler’s evaluation of a schedule is based on the evaluation of individual constraints and of their weighted aggregation, the user has the full control over the scheduling pro- cess with the possibility to experiment with different settings, the scheduler applies iterative improvement methods to optimize solutions, and the framework shall be extendible and refinable. The goal is to support the construction of dedicated scheduling systems for a specific application and not to develop a system capable of scheduling different applica- tions. 1.1 Constraint-based Representation of Schedules Scheduling is an activity controlled by constraints and guided by several objective functions. Usually scheduling is described as a problem of satisfying temporal con- straints. However, temporal constraints as processing times and due dates and such objectives as minimization of the makespan or of the mean flow-time are often insufficient to represent industrial problems. The DÉJÀ VU framework supports further constraints and objectives like compatibil- ity constraints (chemical and format), idle time constraints, minimization of substitutable resources, restricted capac- ity, or equilibre load of sharable resources. These con- straint types have been derived from several scheduling problems in the steel industry. Other constraint types from new domains can be generated by derivation from existing ones with small effort, because we have a very general approach to represent them. Many constraints of industrial production environments are soft and relaxable and moreover they may be contradic- tory and a trade-off between these constraints must be found for a good solution. These requirements are reflected by our constraint model: A constraint is a relation between two or more scheduling objects and/or attributes. This rela- tion is mapped on a satisfaction degree that evaluates how good this constraint is satisfied in the actual schedule. Different constraint types obtain a domain dependent weight reflecting the constraint’s importance for the do- main. A schedule evaluation is done by a weighted aggre- gation of all satisfaction degrees. For each constraint type we can specify further a threshold to decide whether the constraint violation is hard. The constraint’s weight and the threshold can be modified in a scheduling session to experiment with different settings.