Friday, December 11 or Saturday, December 12, 2009
Whistler, B.C., Canada, at the Westin Resort and Spa and the
Hilton Whistler Resort and Spa
Please send an extended abstract of max. 1 page describing the poster you intend to present to
mpd37 (at) cam.ac.uk
Choose a format of your liking, e.g., the standard NIPS template.
The deadline for abstract submissions is October 17, 2009.
The notification will be October 26, 2009.
During the last decade, many areas of Bayesian machine learning have reached a high level of maturity. This has resulted in a variety of theoretically sound and efficient algorithms for learning and inference in the presence of uncertainty. However, in the context of control, robotics, and reinforcement learning, uncertainty has not yet been treated with comparable rigor despite its central role in risk-sensitive control, sensorimotor control, robust control, and cautious control. A consistent treatment of uncertainty is also essential when dealing with stochastic policies, incomplete state information, and exploration strategies.
A typical situation where uncertainty comes into play is when the exact state transition dynamics are unknown and only limited or no expert knowledge is available and/or affordable. One option is to learn a model from data. However, if the model is too far off, this approach can result in arbitrarily bad solutions. This model bias can be sidestepped by the use of flexible model-free methods. The disadvantage of model-free methods is that they do not generalize and
often make less efficient use of data. Therefore, they often need more trials than feasible to solve a problem on a real-world system. A probabilistic model could be used for efficient use of data while alleviating model bias by explicitly representing and incorporating uncertainty.
The use of probabilistic approaches requires (approximate) inference algorithms, where Bayesian machine learning can come into play. Although probabilistic modeling and inference conceptually fit
into this context, they are not widespread in robotics, control, and reinforcement learning. Hence, this workshop aims to bring researchers together to discuss the need, the theoretical properties, and the practical implications of probabilistic methods in control, robotics, and reinforcement learning.
One particular focus will be on probabilistic reinforcement learning approaches that profit recent developments in optimal control which show that the problem can be substantially simplified if certain structure is imposed. The simplifications include linearity of the (Hamilton-Jacobi) Bellman equation. The duality with Bayesian estimation allow for analytical computation of the optimal control laws and closed form expressions of the optimal value functions.
The workshop will consist of short invited presentations and a session with contributed posters (plus poster spotlight). Topics (from a theoretical and practical perspective) to be addressed include, but are not limited to:
– How can we efficiently plan and act in the presence of uncertainty in states/rewards/observations/environment?
– Shall we model the lack of knowledge or can we simply ignore it?
– How can prior knowledge (e.g., expert knowledge and domain knowledge) be incorporated?
– How much manual tuning and human insight (e.g., domain knowledge) is a) required and b) available to achieve good performance?
– Is there a principled way to account for imprecise models and model bias?
– What roles should probabilistic models play in control? Are they needed at all?
– What kinds of probabilistic models are useful?
– In traditional control, hand-crafted control laws often prevail since optimal control laws are mostly too aggressive due to model errors while robust control laws can be too conservative since they always assume the worst case. Can “probabilistic control” bridge the gap between robust and optimal control laws?
– How can we exploit the linearity of the (Hamilton-Jacobi) Bellman equation and the duality with Bayesian estimation?
– Can we compute the optimal control law analytically and is there a closed-form expression of the value function?
– How can existing machine learning methods be applied to efficiently solve stochastic control problems?
Dieter Fox (University of Washington), confirmed
Drew Bagnell (CMU), pending
Evangelos Theodorou (USC), confirmed
Jovan Popovic (MIT), confirmed
Konrad Koerding (Northwestern University), confirmed
Marc Toussaint (TU Berlin), confirmed
Miroslav Karny (Academy of Sciences of the Czech Republic), confirmed
Mohammad Ghavamzadeh (INRIA), pending
Roderick Murray-Smith (University of Glasgow), pending
Bert Kappen (University of Nijmegen), confirmed
Emanuel Todorov (University of Washington), confirmed
Marc Peter Deisenroth
Carl Edward Rasmussen