Postgraduate Course: Reinforcement Learning (INFR11010)
|School||School of Informatics
||College||College of Science and Engineering
||Availability||Available to all students
|Credit level (Normal year taken)||SCQF Level 11 (Postgraduate)
|Home subject area||Informatics
||Other subject area||None
||Taught in Gaelic?||No
|Course description||This module covers a range of adaptive learning systems, in particular reinforcement learning and unsupervised methods, particularly as used in RL systems. By the end of the module the student should have a grasp of modern learning techniques and the issues involved in dealing with real-world data. The main techniques covered in the course are basic reinforcement learning, dynamic programming, Monte Carlo methods, Q-learning, function approximation, unsupervised and constructive methods, radial basis and other local functions, classifier systems as compared to RL systems.
Entry Requirements (not applicable to Visiting Students)
||Co-requisites|| It is RECOMMENDED that students also take
Natural Computing (INFR09038)
||Other requirements|| For Informatics PG and final year MInf students only, or by special permission of the School. Students should be familiar with the mathematical concepts therein, particularly vectors and matrices, partial differentiation, and some probability.
|Additional Costs|| None
Information for Visiting Students
|Displayed in Visiting Students Prospectus?||Yes
Course Delivery Information
|Delivery period: 2012/13 Semester 2, Available to all students (SV1)
||Learn enabled: No
|Central||Lecture||1-11|| 12:10 - 13:00|
|Central||Lecture||1-11|| 12:10 - 13:00|
||Week 1, Tuesday, 12:10 - 13:00, Zone: Central. Lecture Theatre 4, 7 Bristo Square |
|No Exam Information
Summary of Intended Learning Outcomes
|1 - Knowledge of basic and advanced reinforcement learning techniques.
2 - Insight into the problems involved in applying these techniques to deal with real world data, and how to overcome those problems.
3 - Appreciation and identification of suitable learning tasks to which these learning techniques can be applied
4 - Ability to evaluate how effective a particular learning procedure has been -- internal indicators of learning success vs. actual behaviour of the learner.
5 - Use and writing of Matlab programs, ability to set up and run computational experiments to produce statistically sound results
6 - Formulation of problems, evaluation of results from the student's own experiments and those presented in some cases in the research literature.
|Written Examination 80|
Assessed Assignments 20
Oral Presentations 0
Two assignments are set, each accounting for 10% of the overall mark. Typically they require programming a learning system or experimenting with an existing system, using MATLAB.
If delivered in semester 1, this course will have an option for semester 1 only visiting undergraduate students, providing assessment prior to the end of the calendar year.
||The main topics to be covered are some or all of the following (there are some changes from year to year)
* Reinforcement learning framework
* Bandit problems and action selection
* Dynamic programming methods
* Monte-Carlo methods
* Temporal difference methods
* Q-learning and eligibility traces
* Environment modelling
* Function approximation for generalisation
* Actor-critic, applications
* Planning in the RL context
* Unsupervised, self-organising networks and RL
* Constructive methods - nets that grow
* Evaluating performance
Relevant QAA Computing Curriculum Sections: Artificial Intelligence, Data Structures and Algorithms, Intelligent Information Systems Technologies, Simulation and Modelling
||# Reinforcement Learning. An Introduction. Richard S. Sutton and Andrew G. Barto. MIT Press, Cambridge MA, 1998.
# Other material as handouts or on web page
Timetabled Laboratories 0
Non-timetabled assessed assignments 25
Private Study/Other 55
|Course organiser||Dr Iain Murray
Tel: (0131 6)51 9078
|Course secretary||Miss Kate Weston
Tel: (0131 6)50 2701
© Copyright 2012 The University of Edinburgh - 14 January 2013 4:09 am