Site en anglais



Unité : Scheduling Optimisation Security | ULB842

Description :

The goal of the PRALINEH project is to design new autonomous planning optimization algorithms that can be applied to very hard and
dynamic real-life cases, thus able to make optimal decisions in near real-time. By optimal we mean decisions that can improve
significantly operational KPIs such as lead-times, on-time-delivery, inventory, equipment efficiency, throughput and last but not
least use of resources whether these are consumables, energy, waste, … thus contributing to higher durability and lower footprint.
The economic impact of these optimizations can be easily as lower costs and higher revenues. It is common knowledge that legacy
planning & scheduling solutions do not provide the flexibility, agility and optimality required as they are often just
simulations or interactive charts with very basic logic.
To build the system that we want, we need new parallel metaheuristics and real-time optimization tools based upon machine
learning methods. The originality of the project is to combine the following four elements: parallel execution (for higher performance
and lower response latency), real-time constraints (to satisfy use case limits), various metaheuristics (as each problem poses new
challenges) and machine learning (in order to both fine tune the heuristics and to adapt to uncertain or low-quality data)
	Parallel execution: in order to speed up the execution of the heuristics upon modern, i.e., multicore, architectures the
heuristics must be parallelized (multithreading). This poses various synchronization challenges.
	Metaheuristics: we assume the genericity of the target applications: making relatively few assumptions about the
optimization problems. Thus, we need to use a large toolset of metaheuristics and let the system discover which one is the best fit. This
way, the solution may be usable for a variety of problems.
	Real-time constraints: while the project targets generic industry and logistics scheduling problems we target real-time
applications where the time to decide when and at which each scheduled activity completes is important, typically we have to meet a
deadline or at least minimize the latency.
	Machine learning, we aim to use those techniques at two different levels:
o	Tuning the metaheuristic’s parameters (model and solver), e.g., the degree of parallelism (number of threads). This offline
learning stage would be based on Monte Carlo experiments and self-generated sample data.
o	Re-tuning

Liste des bailleurs :

  • Région wallonne