Register now After registration you will be able to apply for this opportunity online.
This opportunity is not published. No applications will be accepted.
Bayesian Dual Iterative Learning Control
In this project, we want to develop algorithms that are capable of simultaneously optimizing and actively learning about a process with the ultimate goal of enabling on demand system identification.
Keywords: Dual Control, Bayesian Inference, Iterative Learning Control
Questions about how to efficiently exploit data to improve the efficiency and robustness of engineered systems are becoming increasingly pertinent in our ever more data rich world. In many situations, we reply on a model of the process during optimization. The sensitivity of this model is critical to predicting which actions are likely to lead to improvements. Unfortunately, these models are imperfect and inaccuracies can stymie progress. In this case it is necessary to gather more data to improve the model in order to continue optimization.
This project aims to develop a variant of iterative learning control (ILC) that actively explores the input space whenever it is necessary to further optimize the process. In ILC, an agent learns how to perform a task through repetition, i.e., by taking an action, observing the result, and adjusting accordingly. For example, a basketball player learning to shoot a free throw. In our case, the goal is to have a CNC machine quickly and accurately cut a specified part geometry despite disturbances such as friction, inertia and so on. Since the goal is to automatically balance the dual goals of optimization and learning (i.e., exploration vs. exploitation) this is an instance of a dual control problem. Our approach will be to combine Bayesian linear regression with sequential optimal experimental design and approximate the resulting problem with MPC.
**Remark** This project is also closely related to the area of optimal design of experiments.
**Remark** While motivated by precision manufacturing, we expect this project will be more focused on dual control methodology development.
Questions about how to efficiently exploit data to improve the efficiency and robustness of engineered systems are becoming increasingly pertinent in our ever more data rich world. In many situations, we reply on a model of the process during optimization. The sensitivity of this model is critical to predicting which actions are likely to lead to improvements. Unfortunately, these models are imperfect and inaccuracies can stymie progress. In this case it is necessary to gather more data to improve the model in order to continue optimization.
This project aims to develop a variant of iterative learning control (ILC) that actively explores the input space whenever it is necessary to further optimize the process. In ILC, an agent learns how to perform a task through repetition, i.e., by taking an action, observing the result, and adjusting accordingly. For example, a basketball player learning to shoot a free throw. In our case, the goal is to have a CNC machine quickly and accurately cut a specified part geometry despite disturbances such as friction, inertia and so on. Since the goal is to automatically balance the dual goals of optimization and learning (i.e., exploration vs. exploitation) this is an instance of a dual control problem. Our approach will be to combine Bayesian linear regression with sequential optimal experimental design and approximate the resulting problem with MPC.
**Remark** This project is also closely related to the area of optimal design of experiments.
**Remark** While motivated by precision manufacturing, we expect this project will be more focused on dual control methodology development.
The goals of the project are as follows:
- Learn about ILC, dual control, and Bayesian estimation;
- Define a meaningful ILC problem including a model and objective etc.;
- Derive Bayesian prediction and parameter update equations;
- Explore various dual controller formulations and demonstrate them in simulation.
**Publications** If the final results are promising they can potentially be turned into a publication.
**Corona Disclaimer** This project can be done in person at the Automatic Control Lab, hybrid, or completely remotely. Most importantly, we can change between these forms as needed.
**Qualifications**
We are looking for a highly motivated student with background in control, probability and statistics.
- No specific experience with manufacturing systems is necessary;
- Some background in probability, statistics (ideally Bayesian) and control is required.
- Some familiarity with optimization would be helpful but is not necessary
- Enrollment in a masters program;
- Proficiency in English.
The goals of the project are as follows:
- Learn about ILC, dual control, and Bayesian estimation;
- Define a meaningful ILC problem including a model and objective etc.;
- Derive Bayesian prediction and parameter update equations;
- Explore various dual controller formulations and demonstrate them in simulation.
**Publications** If the final results are promising they can potentially be turned into a publication.
**Corona Disclaimer** This project can be done in person at the Automatic Control Lab, hybrid, or completely remotely. Most importantly, we can change between these forms as needed.
**Qualifications** We are looking for a highly motivated student with background in control, probability and statistics.
- No specific experience with manufacturing systems is necessary;
- Some background in probability, statistics (ideally Bayesian) and control is required.
- Some familiarity with optimization would be helpful but is not necessary
- Enrollment in a masters program;
- Proficiency in English.
Please send your resume/CV (including lists of relevant publications/projects) and transcript of records in PDF format via email to dliaomc@ethz.ch.
Please send your resume/CV (including lists of relevant publications/projects) and transcript of records in PDF format via email to dliaomc@ethz.ch.