Robust Continuous Machine Learning of Complex Realtime Communication - verkefni lokið

Fréttatilkynning verkefnisstjóra


The importance of automatic dialogue systems using natural language is increasing in all sorts of applications. The variety and diversity of how people use language is large, and programming in all required knowledge a-priori is impossible. Automatic adaptation – machine learning – is an obvious way to address this problem.

 However, limitations inherent in standard machine learning techniques, e.g. reinforcement learning, prevents them from being applied directly to many tasks within dialogue, as well as a vast number of other real-world tasks which require handling of continuous variables, as opposed to discrete.

Heiti verkefnis: Robust Continuous Machine Learning of Complex Realtime Communication
Verkefnisstjóri: Kristinn R. Þórisson, Vitvélastofnun Íslands ses.
Tegund styrks: Verkefnisstyrkur
Styrkár: 2011-2013
Fjárhæð styrks: 19,1 millj. kr. alls
Tilvísunarnúmer Rannís: 110023

We have developed new ways for extending reinforcement learning techniques to handle continuous variables, greatly improving the applicability of such learning to a new range of tasks, especially those for which discretization of variables is not possible. We have also demonstrated how such technology can be used in a broad dialogue system that can adapt to its user's speaking style. Furthermore, going well beyond speaking style and cadence, we have made significant progress on designing machines that automatically learn a vocabulary – and its meaning – by observation. Lastly, we have developed new techniques for automatically evaluating machine learning algorithms, which greatly shortens the time required to produce task variations for testing machine learning algorithms, allows for testing algorithms that work with continuous variables, and makes feasible close comparisons between machine learning algorithms of different kinds.


Bieger, J., Thórisson, K. R., & Garrett, D. (2014). Raising AI: Tutoring Matters. Proceedings of the Seventh Conference on Artificial General Intelligence (pp. 1-10). Quebec City, Canada: Springer.

Garrett, D., Bieger, J., & Thórisson, K. R. (2014). Tunable and Generic Problem Instance Generation for Multi-objective Reinforcement Learning. IEEE Symposium on Adaptive Dynamic Programming and Reinforcement Learning. Orlando, Florida.

E. Nivel, K. R. Thórisson, B. R. Steunebrink.,H. Dindo, G. Peluzo, M. Rodriguez, C. Hernandez, D. Ognibene, J. Schmidhuber, R. Sanz, H. P. Helgason, A. Chella & G. Jonsson (2014). Autonomous Acquisition of Natural Language. A. P. dos Reis,P. Kommers & P. Isaías (eds.), Proceedings of the IADIS International Conference on Intelligent Systems & Agents 2014 (ISA-14), 58-66, Lissbon, Portugal, July 15-17

Helgason, H. P., K. R. Thórisson, E. Nivel & P. Wang (2013). Predictive Generative Heuristics for Decision-Making in Real-World Environments. K-U Kühnberger, S. Rudolph & P. Wang (eds), Proceedings of Artificial General Intelligence (AGI-13), 50-59, Beijing, China

Jonsdottir, G., K. R. Thórisson (2013). A Distributed Architecture for Real- Time Dialogue and On-Task Learning of Efficient Cooperative Turn- Taking. M. Rojc and N. Campbell (eds), Speech, Gaze and Affect, Ch. 12, 293-324. Boca Raton, Florida, US: Taylor & Francis.

D. Garrett, G. Jonsdottir, K. R. Thórisson, H. Pourvatan (2011). An Architecture for Learning Complex Dialog Skills in Human-Computer Spoken Interaction. Icelandic Institute for Intelligent Machines Technical Report, 2011.

Thórisson, K. R., Bieger, J. & Garrett, D. (in preparation). Abstract Task Definitions in WavesWorld: A Tool for Evaluating Artificial Cognitive Development. To be submitted to the Artificial General Intelligence conference 2015.

Garrett, D., Bieger, J. & Thórisson, K. R. (in preparation). Evaluation Framework for Machine Learning Algorithms. To be published in Journal of Artificial Intelligence and Soft Computing Research

Þetta vefsvæði byggir á Eplica