In this paper we present a new approach to plan understanding that explains observed actions in terms of domain knowledge. The process operates over hierarchical methods and utilizes an incremental form of data-driven abductive inference. We report experiments on problems from the Monroe corpus that demonstrate a basic ability to construct plausible explanations, graceful degradation of performance with reduction of the fraction of actions observed, and results with incremental processing that are comparable to batch interpretation. We also discuss research on related tasks such as plan recognition and abductive construction of explanations.

Modern multi-core computers provide an opportunity to parallelize plan recognition algorithms to decrease runtime. Viewing the problem as one of parsing and performing a complete breadth first search, makes ELEXIR (Engine for LEXicalized Intent Recognition) (Geib '09, Geib '11) particularly suitable for such parallelism. This paper documents the extension of ELEXIR to utilize such modern computing platforms. We will discuss multiple possible algorithms for distributing work between parallel threads and the associated performance wins. We will show, that the best of these algorithms will provide close to linear speedup (up to a maximum number of processors), and that features of the problem domain have an impact on the speedup.

There are currently no strong methods for planning in a stochastic domain, with low-level sensors that are limited and possibly inaccurate. Existing architectures have flaws that make their use in a real-world environment impractical. We propose an architecture that utilizes POMDPs to create a hierarchical planning system. This system is capable of developing macro-actions that can expedite planning on a large scale, and can learn new plans quickly and efficiently, without deliberate design by the programmer.

We advocate the position that unsupervised learning of rich representations requires careful consideration of an issue that usually receives only cursory attention: The definition of a statistical ‘event’, or ‘sample’.  Data sets are presumed to have been generated by sampling from some probability distribution that is to be estimated, but there is no general canonical way to select a model for a given data set and define the correspondence between the various components of its joint random variable and particular subsets, or more generally, features, of the data.  Any attempt to automate this choice must confront the fact that without a definition of ‘event’, this exercise cannot be formulated as a statistical learning problem.  We introduce two supplementary criteria, information at a distance and information contrast , in order to clear this impasse, and show anecdotal results from using each.  We argue that this issue also arises (whether recognized or not) in automated learning of feature hierarchies to form a rich representations, because distinct events are selected at one level of the hierarchy and bound together to form joint events at the next level.

We present a top-down approach for learning abstractions whereby a robot begins with a coarse representation of the world and incrementally finds new distinctions as they enable the robot to better predict its environment. The approach has been implemented on a simulated robot that learns new distinctions in the form of variable discretizations through autonomous exploration. This paper discusses how to generalize this approach to learning broader abstractions.

We propose a knowledge-representation architecture allowing a robot to learn arbitrarily complex, hierarchical / symbolic relationships between sensors and actuators. These relationships are encoded in high-dimensional, low-precision vectors that are very robust to noise. Low-dimensional (single-bit) sensor values are projected onto the highdimensional representation space using low-precision random weights, and the appropriate actions are then computed using elementwise vector multiplication in this space. The high-dimensional action representations are then projected back down to low-dimensional actuator signals via a simple vector operation like dot product. As a proof-of-concept for our architecture, we use it to implement a behavior-based controller for a simulated robot with three sensors (touch sensor, left/right light sensor) and two actuators (wheels). We conclude by discussing the prospects for deriving such representations automatically.

Computer vision and artificial intelligence research has long danced around the subject of causality: vision researchers use causal relationships to aid action detection, and AI researchers propose methods for causal induction independent of video sensors. In this paper, we argue that learning perceptual causality from video is a necessary step for understanding scenes in video. We explain how current object and action detection is suffering without causality, and we explain how current causality research is suffering without grounding on raw sensors. We then go on to describe one plausible solution for grounding perceptual causality on raw sensors.

Learning representations from data is one of the fundamental problems of artificial intelligence and machine learning. Many different approaches exist for learning representations, but what constitutes a good representation is not yet well understood. In this work, we view the problem of representation learning as one of learning features (e.g., hidden units of neural networks) such that performance of the underlying base system continually improves. We study an important case where learning is done fully online (i.e., on an example-by-example basis) from an unending stream of data. In the presence of an unending stream of data, the computational cost of the learning element should not grow with time and cannot be much more than that of the performance element. Few methods can be used effectively in this case. We show that a search approach to representation learning can naturally fit with this setting. In this approach good representations are searched by generating different features and then testing them for utility. We develop new representation-search methods and show that the generate-and-test approach can be utilized in a simple and effective way for learning representations. Our methods are fully online and add only a small fraction to the overall computation. They constitute an important step toward effective and inexpensive solutions to representation learning problems.

We consider the problem of how to plan efficiently in low-level, continuous state spaces with temporally abstract actions (or skills), by constructing abstract representations of the problem suitable for task-level planning.The central question this effort poses is which abstract representations are required to express and evaluate plans composed of sequences of skills. We show that classifiers can be used as a symbolic representation system, and that the ability to represent the preconditions and effects of an agent's skills is both necessary and sufficient for task-level planning.The resulting representations allow a reinforcement learning agent to acquire a symbolic representation appropriate for planning from experience.

A human-level artificially intelligent agent must be able to represent and reason about the world, at some level, in terms of high-level concepts such as entities and relations. The problem of acquiring these rich high-level representations, known as the knowledge acquisition bottleneck, has long been an obstacle for achieving human-level AI. A popular approach to this problem is to handcraft these high-level representations, but this has had limited success. An alternate approach is for rich representations to be learned autonomously from low-level sensor data. Potentially, the latter approach may yield more robust representations, and should rely less on human knowledge-engineering. The papers in this workshop present work and strategies in this latter approach.