Conference Agenda

In a time-dynamic setting, job schedules must be continuously reevaluated and possibly adjusted, for instance, in response to changing priorities triggered by significant customer orders. This situation becomes even more challenging in a multi-criteria context where the attainment of multiple conflicting objectives must be managed over a certain time horizon. This is because each scheduling/rescheduling time-step involves calculating multiple efficient solutions, namely optimal schedules, from which one must select a solution for further execution within the planning horizon. In doing so, the selection process should not only be backward-looking but also forward-looking, in order to maintain planning flexibility and to uphold the attainability of desired targets in the future. To facilitate this in a decision-oriented manner, ensuring that overall performance goals are at least approximately achieved by the end of a planning horizon, this presentation discusses the concept of goal-stable programming and its applicability in the context of multi-criteria online scheduling.

This research addresses the intricate dynamics of planning in industrial settings, traditionally segmented into strategic, tactical, and operational levels. These levels have been effective in guiding long-term goals, medium-term objectives, and short-term operational needs. Recently, the emerging introduction of an execution level has revolutionized production planning by enabling real-time adjustments to cope with unforeseen disruptions, such as machine failures and labor shortages. This adaptive approach enhances resource utilization and ensures smoother execution on the shop floor.

Surprisingly, the realm of maintenance planning has yet to embrace this execution level, despite its potential to significantly improve efficiency and responsiveness, especially in the era of AI and predictive analytics. The persistent separation of production and maintenance planning, lacking synchronous operations, results in notable inefficiencies and resource wastage. Aligning these functions could prevent scenarios where maintenance during peak production times leads to costly downtimes and repairs.

In response to these challenges, this research proposes an innovative framework for real-time maintenance planning, integrating AI-driven maintenance planning and resource allocation strategies. This framework is designed to leverage real-time data to make proactive maintenance decisions along production that minimize disruptions and optimize overall production performance. By aligning maintenance with production processes and employing advanced technologies, this framework aims to significantly enhance the approach industries currently take toward maintenance planning, promoting greater operational efficiency and sustainability.

Hexaly Optimizer is a “model and run” mathematical optimization solver based on various exact and heuristic methods. The presentation will introduce the different components of Hexaly Optimizer’s primal heuristics through disjunctive scheduling problems.

We will first show how its modeling formalism can be used to express various academic and industrial scheduling problems using interval and list decision variables. These models are very compact, which enables the solver to handle even large-scale problems.

Detecting non-overlap constraints in the model provides the solver with valuable information, which can be exploited through various scheduling-specific movements implemented in Hexaly Optimizer’s neighborhood search. However, due to the tightness of precedence and non-overlap constraints in good solutions to disjunctive scheduling problems (Job Shop Scheduling Problem, for example), such a small-neighborhood search alone struggles to obtain good performance.

Hexaly Optimizer overcomes this issue by reinforcing its neighborhood search component with a solution repair algorithm based on constraint propagation. When a move renders the solution infeasible, it is gradually repaired, one constraint at a time, by heuristically shifting the variables just enough to repair. To extend the transformation rather than cancel it, and to ensure the procedure is fast, we impose never to backtrack on a previous decision to increase or decrease a variable’s value.