Many challenging real-world control problems require adaptation and learning in the presence of uncertainty. Examples of these challenging domains include aircraft adaptive control under uncertain disturbances [1], [2], multiple-vehicle tracking with space-dependent uncertain dynamics [3], [4], robotic-arm control [5], blimp control [6], [7], mobile robot tracking and localization [8], [9], cart-pole systems and unicycle control [10], gait optimization in legged robots [11] and snake robots [12], and any other system whose dynamics are uncertain and for which limited data are available for model learning. Classical model reference adaptive control (MRAC) [13]-[15] and reinforcement learning (RL) methods [16]-[23] have been developed to address these challenges and rely on parametric adaptive elements or control policies whose number of parameters or features are fixed and determined a priori. One example of such an adaptive model are radial basis function networks (RBFNs), with RBF centers pre-allocated based on expected operating domains [24], [25].

Information value functions based on the Kullback- Leibler divergence have been shown the most effective for planning sensor measurements using greedy strategies. However, this paper shows that the problem of optimally controlling a sensor over a finite time horizon is NP-hard. The new lexicographic approach presented in this paper relies on new information value functions that are shown to be additive and can be optimized efficiently over time. Also, this paper derives a convex approximation of the sensor nonlinear field-of-view that allows for real-time sensor control, and is accompanied by performance guarantees. Extensive results on real pedestrian data show that the lexicographic control algorithm significantly improves modeling and prediction performance when compared to existing algorithms.

Expectation maximization (EM) has recently been shown to be an efficient algorithm for learning finite-state controllers (FSCs) in large decentralized POMDPs (Dec-POMDPs). However, current methods use fixed-size FSCs and often converge to maxima that are far from optimal. This paper considers a variable-size FSC to represent the local policy of each agent. These variable-size FSCs are constructed using a stick-breaking prior, leading to a new framework called \emph{decentralized stick-breaking policy representation} (Dec-SBPR). This approach learns the controller parameters with a variational Bayesian algorithm without having to assume that the Dec-POMDP model is available. The performance of Dec-SBPR is demonstrated on several benchmark problems, showing that the algorithm scales to large problems while outperforming other state-of-the-art methods.