Machine Control through Bayesian reinforcement learning

---

---

MOON will address the application of Reinforcement Learning solutions to provide tool adaptability and the optimization of the control of two different machines: rib machines and surface machines. For that, MOON will not be only focused on process variable optimization, but also on higher lever processes involved in the productivity of these machines. These higher level processes are the path strategy generation for the rib machine use case, and the post-processing time in the case of surface machining. The symbiotic relationship between these two levels of problem assessment will provide dynamic tool threshold definition to the distributed global solution with its consequent robustness and a significant reduction of time typically invested in calibration, quality control or strategy optimization. This will derive in a substantial improvement in terms of process productivity.

To address partial information issues for the application of Reinforcement Learning in the proposed context, a Bayesian approach will be followed for the evaluation of the system state in a Bayesian Reinforcement Learning framework. Such approach will allow the assessment of system robustness and adaptability under changing conditions,giving a method for the quantification of system flexibility.

In order to deploy such Reinforcement Learning in an efficient way, considering sampling ratios in the range of milliseconds, MOON will develop dimensionality reduction methods based on autoencoders. This will be done to compress the signals involved in the process control into a single dynamic state within a reduced latent space.

MOON will also address the generation of synthetic data using Generative approaches in order to enhance the scalability of the solution.

• Sectors: Machinning
• Financing programme: Horizon 2020
• R&D&i: Artificial intelligence, Flexible manufacturing cells, Integrated monitoring and control systems
• Partners: CEA - Commissariat à l’Energie Atomique et aux Energies Alternatives
• Acronym: MOON
• Ambit: International