Machine Learning and Sensorimotor Control (P00853)

? Credit Points : 10 ? SCQF Level : 11 ? Acronym : INF-P-MLSC

Control of complex, compliant, multi degree of freedom sensorimotor systems like humanoid robots or autonomous vehicles have been pushing the limits of traditional control theoretic methods. This course aims at introducing adaptive and learning control as a viable alternative. The course will take the students through various aspects involved in motor planning, control, estimation, prediction and learning with an emphasis on the computational perspective. We will learn about statistical machine learning tools and methodologies particularly geared towards problems of real-time, online learning for sensorimotor control. Issues and possible approaches for multimodal sensor integration, sensorimotor transformations and learning in high dimensions will be discussed. This will be put in context through exposure to topics in human motor control, experimental paradigms and the use of computational methods in understanding biological sensorimotor mechanisms.

Entry Requirements

? Pre-requisites : Introduction to Vision & Robotics or Learning from Data (Level 10/11). Also, Mathematics for Informatics 3 and Mathematics for Informatics 4 (or equivalent). For Informatics PG students only, or by special permission of the School. Students should have a good grounding in mathematics and be comfortable with linear algebra and matrix computations. A basic understanding of control theory is desirable.

? Co-requisites : None, although MSc students on the Intelligent Robotics theme will be taking one of Advanced Vision and Intelligent Autonomous Robotics.

Subject Areas

Home subject area

Intelligent Robotics, (School of Informatics, Schedule O)

Other subject areas

Learning from Data, (School of Informatics, Schedule O)

Delivery Information

? Normal year taken : Postgraduate

? Delivery Period : Not being delivered

? Contact Teaching Time : 2 hour(s) per week for 10 weeks

Summary of Intended Learning Outcomes

Students who successfully complete this course should be able to:
- Describe the components of 'traditional' model based control and critically assess it's limitations in the real-time control of compliant, high dimensional sensorimotor systems.
- Design and evaluate experimental paradigms to test various biological control hypotheses and identify potential ways of reducing control complexity.
- Apply the machine learning tools and algorithms learnt in the module to design an efficient adaptive (learning) control scheme for a given real world control problem.
- Implement a basic 'model-based' control schema with learning in MATLAB/C using dimensionality reduction techniques.
- Carry out benchmark comparisons against the state of the art learning methods and optimization/planning strategies.