Ph. D. Project
Title:
Machine learning in control of complex systems. Application to cable robots.
Dates:
2020/10/01 - 2023/09/30
Supervisor(s):
Other supervisor(s):
Martinez Dominique DR-CNRS (dominique.martinez@univ-lorraine.fr)
Description:
The thesis subject concerns both the modeling and the control of a class of nonlinear dynamical systems using Artificial Intelligence (AI) techniques. The
application part deals with control of cable robots and the development of a real-time simulation code. This project is part of the scientific collaboration
between CRAN and LORIA - Lorraine University.
While the control theory and state estimation has grown considerably over the past seven decades, as is evident from the many works in the literature,
the use of AI techniques for control opens up a relatively unexplored field with many promises.
To date, automatic control techniques uses (implicitly or explicitly) the model of the system to be controlled in order to construct a state estimator, a control
law or to make a diagnosis strategy. For a class of systems, under certain assumptions such as observability and / or controllability, we succeed in proving
the stability of these approaches. Nevertheless, in many cases, even when the model is known, it is difficult to find analytical solutions based solely on the
model.
In a recent work based on the reinforcement learning technique, with specific activation functions, the authors show that there could be solutions to
constrained optimization problems for a class of dynamic systems, unlike classic approach that does not provide a solution.
The other aspect, representing the major interest of AI techniques for the control, concerns the use of learning algorithms from the available data, when the
model is partially known, difficult to model or totally unknown.
The thesis subject is divided into two parts:
- The first one is methodological whose objective is to develop generic approaches, based on machine learning, for the modeling and control of a class of
dynamic systems. In this study, we must focus on the use of neural networks with a particular attention to real time implementation.
- The second is dedicated to the experimental validation of the obtained results on a cable robot recently built in the CRAN / LORIA. The major
advantage of this platform is that it represents a large class of non-linear systems with the additional difficulty of solving an algebraic stress on the cable
tensions. The control law should guarantee a trajectory tracking, unknown in advance, at high speed> 2 m / s and acceleration> 2 g. The robot's eight
actuators can be thought of as agents who must collaborate and develop an operating strategy to accomplish a mission.
References
[1] I. Goodfellow, Y. Bengio, and A. Courville. Deep learning. MIT-Press, 2016.
[2] Y. LeCun, Y. Bengio, G. Hinton. Deep learning. Nature 521, pp. 436-444, 2015.
[3] Farnaz Adib Yaghmaie and David J Braun. Reinforcement learning for a class of continuous-time input constrained optimal control problems.
Automatica, 99 :221- 227, 2019.
application part deals with control of cable robots and the development of a real-time simulation code. This project is part of the scientific collaboration
between CRAN and LORIA - Lorraine University.
While the control theory and state estimation has grown considerably over the past seven decades, as is evident from the many works in the literature,
the use of AI techniques for control opens up a relatively unexplored field with many promises.
To date, automatic control techniques uses (implicitly or explicitly) the model of the system to be controlled in order to construct a state estimator, a control
law or to make a diagnosis strategy. For a class of systems, under certain assumptions such as observability and / or controllability, we succeed in proving
the stability of these approaches. Nevertheless, in many cases, even when the model is known, it is difficult to find analytical solutions based solely on the
model.
In a recent work based on the reinforcement learning technique, with specific activation functions, the authors show that there could be solutions to
constrained optimization problems for a class of dynamic systems, unlike classic approach that does not provide a solution.
The other aspect, representing the major interest of AI techniques for the control, concerns the use of learning algorithms from the available data, when the
model is partially known, difficult to model or totally unknown.
The thesis subject is divided into two parts:
- The first one is methodological whose objective is to develop generic approaches, based on machine learning, for the modeling and control of a class of
dynamic systems. In this study, we must focus on the use of neural networks with a particular attention to real time implementation.
- The second is dedicated to the experimental validation of the obtained results on a cable robot recently built in the CRAN / LORIA. The major
advantage of this platform is that it represents a large class of non-linear systems with the additional difficulty of solving an algebraic stress on the cable
tensions. The control law should guarantee a trajectory tracking, unknown in advance, at high speed> 2 m / s and acceleration> 2 g. The robot's eight
actuators can be thought of as agents who must collaborate and develop an operating strategy to accomplish a mission.
References
[1] I. Goodfellow, Y. Bengio, and A. Courville. Deep learning. MIT-Press, 2016.
[2] Y. LeCun, Y. Bengio, G. Hinton. Deep learning. Nature 521, pp. 436-444, 2015.
[3] Farnaz Adib Yaghmaie and David J Braun. Reinforcement learning for a class of continuous-time input constrained optimal control problems.
Automatica, 99 :221- 227, 2019.
Keywords:
Non linear systems - AI based Control design - Cable robots
Department(s):
| Control Identification Diagnosis |
