This paper describes the development and empirical testing of an intelligent tutoring system (ITS) with two emerging methodologies: (1) a partially observable Markov decision process (POMDP) for representing the learner model and (2) inquiry modeling, which informs the learner model with questions learners ask during instruction. POMDPs have been successfully applied to non-ITS domains but, until recently, have seemed intractable for large-scale intelligent tutoring challenges. New, ITS-specific representations leverage common regularities in intelligent tutoring to make a POMDP practical as a learner model. Inquiry modeling is a novel paradigm for informing learner models by observing rich features of learners’ help requests such as categorical content, context, and timing. The experiment described in this paper demonstrates that inquiry modeling and planning with POMDPs can yield significant and substantive learning improvements in a realistic, scenario-based training task.

We present the domain-independent HRFF algorithm, which solves goal-oriented HMDPs by incrementally aggregating plans generated by the Metric-FF planner into a policy defined over discrete and continuous state variables. HRFF takes into account non-monotonic state variables, and complex combinations of many discrete and continuous probability distributions. We introduce new data structures and algorithmic paradigms to deal with continuous state spaces: hybrid hierarchical hash tables, domain determinization based on dynamic domain sampling or on static computation of probability distributions' modes, optimization settings under Metric-FF based on plan probability and length. We compare with HAO* on the Rover domain and show that HRFF outperforms HAO* by many order of magnitudes in terms of computation time and memory usage. We also experiment challenging and combinatorial HMDP versions of benchmarks from numeric classical planning, with continuous dead-ends and non-monotonic continuous state variables.

For nearly 25 years my research group has investigated the use of domain knowledge, expressed in some version of mathematical logic, that is refined or exploited by numeric-based learning algorithms. These include what we called knowledge-based neural networks and knowledge-based support vector machines. I will cover the key ideas of these methods, as well as the behind-the-scenes motivations that lead to them. I will also describe why we switched from using the phrase 'prior knowledge' to using 'advice.' Finally, I will cover some of our recent work on fast learning and inference for Markov Logic Networks (which can be viewed as a knowledge-based graphical model).

Non intrusive activity recognition systems typically read values from sensors deployed in an environment and combine them with user annotated activities to build a probabilistic model. Recently, features constructed from activity specific conjunctions of binary sensor values have been shown to improve the classification accuracy. Such systems employ greedy feature induction techniques to find the observation features and combine them with state transition distribution in a Hidden Markov Model or a Conditional Random Field. An exhaustive search for optimum features is infeasible in this exponential feature space. We have recently extended the rule ensemble learning using hierarchical kernels (RELHKL) framework, that learns a sparse set of simple features and their optimum weights, to structured output spaces for learning optimum observation features along with the transition features and their weights. The exponentially large space of conjunctions is handled efficiently by exploiting its hierarchical structure. Our experiments have shown good improvement over other approaches. Although such approaches solve propositional classification problems optimally, their first-order extension is non-trivial and is a challenging problem. In this paper, we discuss about the challenges involved in leveraging the RELHKL in structured output spaces approach to learn optimum features in complex first order activity recognition settings.

Recognizing social interests plays an important role of aiding human-computer interaction and human collaborative works. The recognition of social interest could be of great help to determine the smoothness of the interaction, which could be an indicator for group work performance and relationship. From socio-psychological theories, social engagement is the observable form of inner social interest, and represented as patterns of turn-taking and speech emotion during a face-to-face conversation. With these two kinds of features, a multi-layer learning structure is proposed to model the continuous trend of engagement. The level of engagement is classified into “high” and “low” two levels according to human-annotated score. In the result of assessing two-level engagemet, the highest accuracy of our model can reach 79.1%.

A major challenge of pervasive context-aware computing and intelligent environments resides in the acquisition and modelling of rich and heterogeneous context data. Decisive aspects of this information are the ongoing human activities at different degrees of granularity. We conjecture that ontology-based activity models are key to support interoperable multilevel activity representation and recognition. In this paper, we report on an initial investigation about the application of probabilistic description logics (DLs) to a framework for the recognition of multilevel activities in intelligent environments. In particular, being based on Log-linear DLs, our approach leverages the potential of highly expressive description logics with probabilistic reasoning in one unified framework. While we believe that this approach is very promising, our preliminary investigation suggests that challenging research issues remain open, including extensive support for temporal reasoning, and optimizations to reduce the computational cost.

Finding a meaningful way of characterizing the difficulty of partially observable Markov decision processes (POMDPs) is a core theoretical problem in POMDP research. State-space size is often used as a proxy for POMDP difficulty, but it is a weak metric at best. Existing work has shown that the covering number for the reachable belief space, which is a set of belief points that are reachable from the initial belief point, has interesting links with the complexity of POMDP planning, theoretically. In this paper, we present empirical evidence that the covering number for the reachable belief space (or just ``covering number", for brevity) is a far better complexity measure than the state-space size for both planning and learning POMDPs on several small-scale benchmark problems. We connect the covering number to the complexity of learning POMDPs by proposing a provably convergent learning algorithm for POMDPs without reset given knowledge of the covering number.

Many first-order probabilistic models can be represented much more compactly using aggregation operations such as counting. While traditional statistical relational representations share factors across sets of interchangeable random variables, representations that explicitly model aggregations also exploit interchangeability of random variables within factors. This is especially useful in decision making settings, where an agent might need to reason about counts of the different types of objects it interacts with. Previous work on counting formulas in statistical relational representations has mostly focused on the problem of exact inference on an existing model. The problem of learning such models is largely unexplored. In this paper, we introduce Counting Markov Logic Networks (C-MLNs), an extension of Markov logic networks that can compactly represent complex counting formulas. We present a structure learning algorithm for C-MLNs; we apply this algorithm to the novel problem of generalizing natural language instructions, and to relational reinforcement learning in the Crossblock domain, in which standard MLN learning algorithms fail to find any useful structure. The C-MLN policies learned from natural language instructions are compact and intuitive, and, despite requiring no instructions on test games, win 20% more Crossblock games than a state-of-the-art algorithm for following natural language instructions.

To ensure quality results from unreliable crowdsourced workers, task designers often construct complex workflows and aggregate worker responses from redundant runs. Frequently, they create several alternative workflows to accomplish the task, and choose a single workflow to deploy (perhaps the one that achieves the best performance during early experiments). However, this seemingly natural design paradigm does not achieve the full potential of crowdsourcing. In particular, using a single workflow (even the best) to accomplish a task is suboptimal. We show that alternative workflows can compose synergistically to yield a much higher quality output. We formalize the insight with a novel probabilistic graphical model, design and implement AgentHunt, a POMDP-based controller that dynamically switches between these workflows to achieve higher returns on investment, and design offline and online methods for learning model parameters. Live experiments on Amazon Mechanical Turk demonstrate the superiority of AgentHunt for the practical task of generating NLP training data, yielding up to 50% error reduction and greater net utility compared to previous methods.

In the field of probabilistic robotics, a central problem is to determine a robot's state given knowledge of a time series of control commands and sensor readings. The effects of control commands and the behavior of sensor devices are both modeled probabilistically. A variety of methods are available for deriving the robot's belief state, which is a probabilistic representation of the robot's true state (which cannot be directly known). This paper presents a series of five assignments to teach this material at the advanced undergraduate/graduate level. The theoretical aspect of the work is reinforced by practical implementation exercises using ROS (Robot Operating System), and the Bilibot, an educational robot platform.