Skip to Main content Skip to Navigation
Theses

Lifelong Learning by Endogenous Feedback, application to a Robotic System

Bruno Dato 1
1 IRIT-SMAC - Systèmes Multi-Agents Coopératifs
IRIT - Institut de recherche en informatique de Toulouse
Abstract : Robotic applications are linked to the dynamic sociotechnical environment in which they are embedded. In this scope, self-adaptation is a central concern and the design of intelligent applications in such environments requires to consider them as complex systems. The field of robotics is very broad. The focus is made on systems that adapt to the constraints of their environment and not on mechanics or signal processing. In light of this context, the objective of this thesis is the design of a learning mechanism capable of continuous learning using endogenous feedback (i.e. internal interactions) in dynamic sociotechnical environments. This learning mechanism must also verify several properties that are essential in this context such as: agnosticity, lifelong learning, online learning, self-observation, knowledge generalization, scalability, data volume tolerance and explainability. The main contributions consist of the construction of Endogenous Context Learning and the design of the learning mechanism ELLSA for Endogenous Lifelong Learner by Self-Adaptation. The proposed learning mechanism is based on Adaptive Multi-Agent Systems combined with Context Learning. The creation of Endogenous Context Learning is motivated by the characterization of learning inaccuracies that are detected by local negotiations between agents. Endogenous Context Learning also includes an artificial data generation mechanism to improve learning models while reducing the amount of the required learning data. In a Lifelong Learning setting, ELLSA enables dynamic updating of learning models. It introduces Active Learning and Self-Learning strategies to resolve learning inaccuracies. The use of these strategies depends on the availability of learning data. In order to evaluate its contributions, this mechanism is applied to the learning of mathematical functions and to a real problem in the field of robotics: the Inverse Kinematics problem. The application scenario is the learning of the control of multi-jointed robotic arms. The conducted experiments show that Endogenous Context Learning enables to improve the learning performances thanks to internal mechanisms. They also highlight the properties of the system according to the objectives of the thesis: endogenous feedback, agnosticity, lifelong learning, online learning, self-observation, knowledge generalization, scalability, data volume tolerance and explainability.
Complete list of metadata

https://hal.archives-ouvertes.fr/tel-03341997
Contributor : Bruno Dato <>
Submitted on : Monday, September 13, 2021 - 10:33:16 AM
Last modification on : Tuesday, September 21, 2021 - 3:38:18 AM

File

Apprentissage permanent par fe...
Files produced by the author(s)

Identifiers

  • HAL Id : tel-03341997, version 1

Citation

Bruno Dato. Lifelong Learning by Endogenous Feedback, application to a Robotic System. Multiagent Systems [cs.MA]. Université Toulouse 3 - Paul Sabatier, 2021. English. ⟨tel-03341997⟩

Share

Metrics

Record views

51

Files downloads

18