Contact: Marco Wiering

Machine Learning research studies how knowledge can be learned from observations or experiences of an agent. It can be contrasted to programming an agent --- we do not have to provide all knowledge to an agent, which may be infeasible in case the programmer or designer has incomplete knowledge about an environment. Instead, by learning the necessary knowledge we provide the agent with an additional degree of autonomy --- an agent's behavior is completely determined by its own experiences.

PUBLICATIONS Publications on machine learning

Machine learning algorithms are used for different applications. The purpose of machine learning algorithms is to use observations (experiences, data, patterns) to improve a performance element, which determines how the agent reacts when it is given particular inputs. The performance element may be a simple classifier trying to classify an input instance into a set of categories or it may be a complete agent acting in an unknown environment. By receiving feedback on the performance, the learning algorithm adapts the performance element to enhance its capabilities.

Dependening on the feedback we can distinguish between the following forms of learning: supervised learning, unsupervised learning, and reinforcement learning. In supervised learning, the learning algorithms receives inputs and the correct outputs, and searches for a function which approximates the unknown target function. In unsupervised learning, the agent receives only input data and uses an objective function (such as a distance function) to extract clusters in the input data or particular features which are useful for describing the data. In reinforcement learning, the agent receives an input and an evaluation (reward) of the action selected by the agent, and the learning algorithm has to learn a policy which maps inputs to actions resulting in the best performance.

Research themes

Supervised Learning Learning neural networks for game evaluation functions or approximating continuous, noisy data. Learning decision trees for classifying large data sets. Training hidden Markov models (HMMs) for speech recognition and classifying data streams. Training k-Nearest Neighbours or self-organizing maps for data requiring a spatial structure. Using ensemble methods combining multiple classifiers or predictors to improve the accuracy of single classifiers.

Reinforcement Learning Learning agent policies in simulated multi-agent environments Using RL for navigating spiders on the web. Using RL for learning robot soccer strategies. Using RL for complex problem solving such as network routing, or elevator control.

Ongoing projects