Modulcode:	infAuLearn-01a
Englische Bezeichnung:	Autonomous Learning
Modulverantwortliche(r):	Prof. Dr.-Ing. Sven Tomforde
Turnus:	unregelmäßig (SS20 SS21 SS22 SS23)
Präsenzzeiten:	2V 2Ü 2PÜ
ECTS:	8
Workload:	30 h lectures, 30 h exercises, 30 h practical exercises, 150 h self studies
Dauer:	ein Semester
Modulkategorien:	BSc-Inf-WP (BSc Inf (21)) WI (BSc Inf (15)) MSc-Inf-WP (MSc Inf (21)) 2F-MEd-Inf-WP (MEd-Hdl Inf (21)) 2F-MA-Inf-WP (2F-MA Inf (21)) MSc-WInf-WP-Inf (MSc WInf (21)) WI (MSc Inf (15))
Lehrsprache:	Englisch
Voraussetzungen:	Inf-Math-A Inf-Math-B Inf-Math-C infEInf-01a

Kurzfassung:

Autonomous Learning research aims at understanding how autonomous systems can efficiently learn from the interaction with the environment, especially by having an integrated approach to decision making and learning, allowing systems to autonomously decide on actions, representations, hyper-parameters and model structures for the purpose of efficient learning. The term "autonomous" refers to the ability of the system to learn without or with only very limited external support, which includes manual intervention of humans, availability of pre-defined models or expert knowledge, and availability of large sets of sample data. Specific research topics are: Adaptation of the learning models / techniques based on observations, learning from interaction with the environment, re-using knowledge from one domain in another domain, detection of behaviour that deviates from 'usual' or expected behaviour, and learning from and with other systems of the same kind. The lecture gives an introduction to the field of autonomous learning with a particular focus on a utilisation of the different techniques within intelligent systems. Autonomous Learning is cutting edge research, which means that parts of the lecture are based on current research articles rather than on textbooks. Furthermore, a practitioner's perspective is combined with theoretical understanding of the concepts: the lecture units are combined with traditional exercises but also with practical tasks that have to be solved by making use of techniques discussed in the lecture.

Lernziele:

The overall goal of the course is to derive a basic understanding of the motivation, the general concept, and particularly important methods covering the most prominent parts of the field of autonomous learning. This includes techniques for the following aspects of machine learning:

• Fully autonomous learning behaviour: hyper-parameter optimisation, transfer learning, (self-)evaluation,
• self-awareness or environment-awareness with a major focus on anomaly/novelty detection
• By interaction with the environment via sensors and actuators: reinforcement learning
• By efficiently integrating humans into the learning process: active learning
• By interacting with other intelligent systems: collaborative learning
• By using all the above: meta-learning

Particular goals are:

a) Knowledge / Skills:

Understanding of methods for achieving "intelligence" in technical systems, control of learning behaviour with minimal user interaction, continuous self-improvement of system behaviour, cooperation in learning between distributed technical systems

b) Abilities:

Selection and application of techniques of machine learning in technical systems under real-world conditions to control autonomous system behaviour

c) Competencies:

Ability to analyse autonomous learning processes and their behaviour, to determine and interpret relevant assessment parameters / Competence to plan, design and develop intelligent technical systems with the ability to learn autonomously

Lehrinhalte:

a) Introduction and organisation

b) Machine learning basics

c) Reinforcement Learning

d) Anomaly/Novelty Detection

e) Active Learning

Further (depending on time and scope), the following topics may be covered as well:

f) Transfer Learning

g) Model selection (hyper-parameter optimisation and evaluation)

h) Collaborative Learning

i) Summary and outlook (incl. meta-learning)

Weitere Voraussetzungen:

No mandatory modules. However, basic modules such as mathematics 1 to 3 are considered to be common knowledge. We recommend to visit the "Intelligent Systems" or "Computational Intelligence" modules before. We further assume that students have appropriate Python skills for the practical tasks.

Prüfungsleistung:

The examination consists of several parts:

• In the course of the semester, quizzes on the lecture and exercise contents are to be completed. The quizzes will be graded.
• In the course of the semester, four so-called assignments are to be completed as a group. These are a sequential application and evaluation of (deep) reinforcement techniques in the StarCraft environment used in the practical part. The results are to be presented during the course and will be assessed.
• An elaboration on a selected advanced topic is to be prepared and evaluated in relation to the presented lecture contents. The paper is presented in a short talk at the end of the lecture period. The paper and the presentation will also be assessed.

The overall grade is made up of the parts as follows: 20 % quizzes, 30 % assignments, 10 % presentation, 40 % elaboration.

Lehr- und Lernmethoden:

The lecture is accompanied by an exercise that requires active participation of the students. Task sheets and presence work will be discussed in small groups of students. In addition, a second type of exercise is used where students have to apply the techniques in a StarCraft environment to incrementally develop an AI.

a) The lecture is based on the use of the following teaching methods:

• Explanatory lecture with slide sets
• Blackboard with illustrating examples
• Use of video sequences for illustration purposes
• Interactive elements (questions, tasks)
• Internet-based quiz

b) The exercise uses the following building blocks:

• Exercise sheets with written tasks (e.g. calculations)
• Implementation of individual tasks (mostly in Python using Jupyter Notebooks)
• Reading and assessing scientific publications (i.e. journal articles or conference papers)
• Presentation of the results by participants
• Discussions

c) The practical exercise tasks use the following building blocks: - Practical tasks (e.g., given in Python and a StarCraft framework) need to be solved by actually applying, testing, and analysing the behaviour of techniques from the field of autonomous learning - In particular, 4 assignments allowing for an increasing complexity in 'self-learning systems' have to be solved by the students.

Verwendbarkeit:

Literatur:

Basic literature:

• Thomas Mitchell: Machine Learning, The McGraw-Hill Companies, 1997, ISBN 978-0071154673
• Ethem Alpaydin: Introduction to Machine Learning (Adaptive Computation and Machine Learning). The Mit Press, 3rd revised edition, 2014. ISBN: 978-0262028189
• C. Müller-Schloer, S. Tomforde: Organic Computing - Technical Systems for Survival in the real World
• B. Settles: Active Learning
• M. Yamada, Jianhuii Chen, Yi Chang: Transfer Learning: Algorithms and Applications
• C. Bishop: Pattern Recognition and Machine Learning (Information Science and Statistics)
• Sutton, Richard S., and Andrew G. Barto. Introduction to reinforcement learning. Vol. 135. Cambridge: MIT press, 1998.
• S. Russell and P. Norvig: Künstliche Intelligenz. Ein moderner Ansatz. 3. Aufl. PEARSON

Verweise:

Kommentar: