Past approaches for solving MDPs have several weaknesses: 1) Decision-theoretic computation over the state space can yield optimal results but scales poorly. 2) Value-function approximation typically requires human-specified basis functions and has not been shown successful on nominal ("discrete") domains such as those in the ICAPS planning competitions. 3) Replanning by applying a classical planner to a determinized domain model can generate approximate policies for very large problems but has trouble handling probabilistic subtlety.  This paper presents ReTrASE, a novel MDP solver, which combines decision theory, function approximation and classical planning in a new way. ReTrASE uses classical planning to create basis functions for value-function approximation and applies expected-utility analysis to this compact space. Our algorithm is memory-efficient and fast (due to its compact, approximate representation), returns high-quality solutions (due to the decision-theoretic framework) and does not require additional knowledge from domain engineers (since we apply classical planning to automatically construct the basis functions). Experiments demonstrate that ReTrASE outperforms winners from the past three probabilistic-planning competitions on many hard problems.

Statistical-relational reasoning has received much attention due to its ability to robustly model complex relationships. A key challenge is tractable inference, especially in domains involving many objects, due to the combinatorics involved. One can accelerate inference by using approximation techniques, lazy algorithms, etc. We consider Markov Logic Networks (MLNs), which involve counting how often logical formulae are satisfied. We propose a preprocessing algorithm that can substantially reduce the effective size of MLNs by rapidly counting how often the evidence satisfies each formula, regardless of the truth values of the query literals. This is a general preprocessing method that loses no information and can be used for any MLN inference algorithm. We evaluate our algorithm empirically in three real-world domains, greatly reducing the work needed during subsequent inference. Such reduction might even allow exact inference to be performed when sampling methods would be otherwise necessary.

Boltzmann Machines are a powerful class of undirected graphical models. Originally proposed as artificial neural networks, they can be regarded as a type of Markov Random Field in which the connection weights between nodes are symmetric and learned from data. They are also closely related to recent models such as Markov logic networks and Conditional Random Fields. A major challenge for Boltzmann machines (as well as other graphical models) is speeding up learning for large-scale problems. The heart of the problem lies in efficiently and effectively approximating the partition function. In this paper, we propose a new efficient learning algorithm for Boltzmann machines that allows them to be applied to problems with large numbers of random variables. We introduce a new large-margin variational approximation to the partition function that allows Boltzmann machines to be trained using a support vector machine (SVM) style learning algorithm. For discriminative learning tasks, these large margin Boltzmann machines provide an alternative approach to structural SVMs. We show that these machines have low sample complexity and derive a generalization bound. Our results demonstrate that on multi-label classification problems, large margin Boltzmann machines achieve orders of magnitude faster performance than structural SVMs and also outperform structural SVMs on problems with large numbers of labels.

Exploration for robotic mapping is typically handled using greedy entropy reduction. Here we show how to apply information lookahead planning to a challenging instance of this problem in which an Autonomous Underwater Vehicle (AUV) maps hydrothermal vents. Given a simulation of vent behaviour we derive an observation function to turn the planning for mapping problem into a POMDP. We test a variety of information state MDP algorithms against greedy, systematic and reactive search strategies. We show that directly rewarding the AUV for visiting vents induces effective mapping strategies. We evaluate the algorithms in simulation and show that our information lookahead method outperforms the others.

We propose a new topic model for tracking time-varying consumer purchase behavior, in which consumer interests and item trends change over time. The proposed model can adaptively track changes in interests and trends based on current purchase logs and previously estimated interests and trends. The online nature of the proposed method means we do not need to store past data for current inferences and so we can considerably reduce the computational cost and the memory requirement. We use real purchase logs to demonstrate the effectiveness of the proposed method in terms of the prediction accuracy of purchase behavior and the computational cost of the inference.

Automatic recognition of human activities is among the key capabilities of many intelligent systems with vision/perception. Most existing approaches to this problem require sophisticated feature extraction before classification can be performed. This paper presents a novel approach for human action recognition using only simple low-level visual features: motion captured from direct frame differencing. A codebook of key poses is first created from the training data through unsupervised clustering. Videos of actions are then coded as sequences of super-frames, defined as the key poses augmented with discriminative attributes. A weighted-sequence distance is proposed for comparing two super-frame sequences, which is further wrapped as a kernel embedded in a SVM classifier for the final classification. Compared with conventional methods, our approach provides a flexible non-parametric sequential structure with a corresponding distance measure for human action representation and classification without requiring complex feature extraction. The effectiveness of our approach is demonstrated with the widely-used KTH human activity dataset, for which the proposed method outperforms the existing state-of-the-art.

Many robotic path planning applications, such as search and rescue, involve uncertain environments with complex dynamics that can be only partially observed. When selecting the best subset of observation locations subject to constrained resources (such as limited time or battery capacity) it is an important problem to trade off exploration (gathering information about the environment) and exploitation (using the current knowledge about the environment most effectively) for efficiently observing these environments. Even the nonadaptive setting, where paths are planned before observations are made, is NP-hard, and has been subject to much research. In this paper, we present a novel approach to adaptive informative path planning that addresses this exploration-exploitation tradeoff. Our approach is nonmyopic, i.e. it plans ahead for possible observations that can be made in the future. We quantify the benefit of exploration through the “adaptivity gap” between an adaptive and a nonadaptive algorithm in terms of the uncertainty in the environment. Exploiting the submodularity (a diminishing returns property) and locality properties of the objective function, we develop an algorithm that performs provably near-optimally in settings where the adaptivity gap is small. In case of large gap, we use an objective function that simultaneously optimizes paths for exploration and exploitation. We also provide an algorithm to extend any single robot algorithm for adaptive informative path planning to the multi robot setting while approximately preserving the theoretical guarantee of the single robot algorithm. We extensively evaluate our approach on a search and rescue domain and a scientific monitoring problem using a real robotic system.

This paper deals with Decision-Making in the context of multiattribute utility theory and, more precisely, with the problem of efficiently determining the best alternative w.r.t. an agent's preferences (choice problem). We assume that alternatives are elements of a product set of attributes and that the agent's preferences are represented by a generalized additive decomposable (GAI) utility on this set. Such a function allows an efficient representation of interactions between attributes while preserving some decomposability of the model. GAI utilities can be compiled into graphical structures called GAI networks that can be exploited to solve choice problems using collect/distribute schemes essentially similar to those used in Bayesian networks. In this paper, rather than directly using this scheme on the GAI network for determining the most preferred alternative, we propose to work with another GAI function, acting as an upper-bound on utility values and enhancing the model's decomposability. This method still provides the exact optimal solution but speeds up significantly the search. It proves to be particularly useful when dealing with choice and ranking under constraints and within collective Decision-Making, where GAI nets tend to have a large size. We present an efficient algorithm for determining this new GAI function and provide experimental results highlighting the practical efficiency of our procedure.

Inverse reinforcement learning (IRL) is the problem of recovering the underlying reward function from the behaviour of an expert. Most of the existing algorithms for IRL assume that the expert's environment is modeled as a Markov decision process (MDP), although they should be able to handle partially observable settings in order to widen the applicability to more realistic scenarios. In this paper, we present an extension of the classical IRL algorithm by Ng and Russell to partially observable environments. We discuss technical issues and challenges, and present the experimental results on some of the benchmark partially observable domains.

Recognition of chatting activities in social interactions is useful for constructing human social networks. However, the existence of multiple people involved in multiple dialogues presents special challenges. To model the conversational dynamics of concurrent chatting behaviors, this paper advocates Factorial Conditional Random Fields (FCRFs) as a model to accommodate co-temporal relationships among multiple activity states. In addition, to avoid the use of inefficient Loopy Belief Propagation (LBP) algorithm, we propose using Iterative Classification Algorithm (ICA) as the inference method for FCRFs. We designed experiments to compare our FCRFs model with two dynamic probabilistic models, Parallel Condition Random Fields (PCRFs) and Hidden Markov Models (HMMs), in learning and decoding based on auditory data. The experimental results show that FCRFs outperform PCRFs and HMM-like models. We also discover that FCRFs using the ICA inference approach not only improves the recognition accuracy but also takes significantly less time than the LBP inference method.