Daniel G. Shapiro

Experimental investigations of uncertainty in a rule-based system for information retrieval

Abstract This paper describes the results of an experimental investigation of the effects of different representations of uncertainty in an interactive rule-based expert system for Information Retrieval. We draw on Fuzzy Set Theory both to define the various representations and to help analyse the results. We conclude that specification of an uncertainty calculus is a subtle problem that interacts in several ways with the scheme used to represent the expert knowledge itself. Our results indicate that some calculi appear to prevent the development of good queries while those whose behaviour is appropriately smooth can give satisfactory performance. More interestingly, our evidence suggests that as queries become more complex the impact of the choice of calculus is reduced. The paper concludes with a discussion of the insights gained with respect to the general problem of building rule-based expert systems.

Using background knowledge to speed reinforcement learning in physical agents

This paper describes Icarus, an agent architecture that embeds a hierarchical reinforcement learning algorithm within a language for specifying agent behavior. An Icarus program expresses an approximately correct theory about how to behave with options at varying levels of detail, while the Icarus agent determines the best options by learning from experience. We describe Icarus and its learning algorithm, then report on two experiments in a vehicle control domain. The first examines the benefit of new distinctions about state, whereas the second explores the impact of added plan structure. We show that background knowledge increases learning rate and asymptotic performance, and decreases plan size by three orders of magnitude, relative to the typical formulation of the learning problem in our test domain.

An Application of Transfer to American Football: From Observation of Raw Video to Control in a Simulated Environment

Automatic transfer of learned knowledge from one task or domain to another offers great potential to simplify and expedite the construction and deployment of intelligent systems. In practice however, there are many barriers to achieving this goal. In this article, we present a prototype system for the real-world context of transferring knowledge of American football from video observation to control in a game simulator. We trace an example play from the raw video through execution and adaptation in the simulator, highlighting the systems component algorithms along with issues of complexity, generality, and scale. We then conclude with a discussion of the implications of this work for other applications, along with several possible improvements.

Controlling physical agents through reactive logic programming

This paper describes Icarus, a language for specifying the behavior of agents that operate in physical domains. This language provides a novel metaphor of “reactive logic programming”, which makes it convenient to express both extremely reactive control programs and programs with non-trivial deliberative elements. The key features of Icarus are the ability to express hierarchical objectives, requirements, and actions, the use of Prolog-like semantics across function calls, a merged concept of state and action, and a sequence primitive, all embedded in a reactive control loop that considers every relevant action on every cycle of the interpreter. We use a body of examples to illustrate these features, and justify several claims about the expressivity of Icarus relative to existing reactive languages.

Designing Embedded Agents to Optimize End-User Objectives

We formulate the design of discrete-state stochastic control systems as optimizing a performance objective specified in user-oriented terms, i.e. terms that need not be perceivable by the controller or agent being designed. This addresses a user acceptance issue: while agent designs (control algorithms) are limited to distinctions about state supported by artificial perception systems, end users want to evaluate performance using terms such as safety, opportunity, and throughput. We elucidate a feedback from evaluation to agent design via a sensitivity analysis, obtaining the gradient of a time-averaging objective function w.r.t. state transitions influenced by the agent. This gradient leads to a methodology for iteratively improving a systems performance, as perceived by others.

Advancing AI Research and Applications by Learning from What Went Wrong and Why

This special issue of AI Magazine is dedicated to the proposition that problems populate the path to insight, implying the experiences and lessons learned should be shared.

ARTIFICIAL INTELLIGENCE METHODS FOR AMBIENT INTELLIGENCE

Ambient intelligence is a term coined in Europe at the turn of the century to identify the methodologies and technologies that enable an environment to better respond to a user’s needs. The core concept is to endow an environment with the computational power sufficient to sense its inhabitants and to interpret their actions and interactions in order to anticipate their needs, supply them with necessary information, and/or to act on their behalves. Ambient intelligence covers a wide spectrum of applications, ranging from entertainment services to safety and security. This special issue introduces the concept of intelligent environments, and explores the algorithms required to build such systems. The articles address a range of applications but emphasize domestic contexts (i.e., smart homes), which have been a major motivator of ambient intelligence research. Taken together, the articles provide a window into the technologies most relevant to this area of research, emphasizing agent-based methods for information fusion, situation recognition, planning, monitoring, and behavior modeling. The issue contains six papers, which we summarize briefly, below:

Learning Hierarchical Skills from Observation

This paper addresses the problem of learning control skills from observation. In particular, we show how to infer a hierarchical, reactive program that reproduces and explains the observed actions of other agents, specifically the elements that are shared across multiple individuals. We infer these programs using a three-stage process that learns flat unordered rules, combines these rules into a classification hierarchy, and finally translates this structure into a hierarchical reactive program. The resulting program is concise and easy to understand, making it possible to view program induction as a practical technique for knowledge acquisition.

Revising Engineering Models: Combining Computational Discovery with Knowledge

Developing mathematical models that represent physical devices is a difficult and time consuming task. In this paper, we present a hybrid approach to modeling that combines machine learning methods with knowledge from a human domain expert. Specifically, we propose a system for automatically revising an initial model provided by an expert with an equation discovery program that is tightly constrained by domain knowledge. We apply our system to learning an improved model of a battery on the International Space Station from telemetry data. Our results suggest that this hybrid approach can reduce model development time and improve model quality.

The Special Issue of AI Magazine on Structured Knowledge Transfer

This issue summarizes the state of the art in structured knowledge transfer, which is an emerging approach to the general problem of knowledge acquisition and reuse. Its goal is to capture, in a general form, the internal structure of the objects, relations, strategies, and processes used to solve tasks drawn from a source domain, and exploit that knowledge to improve performance in a target domain.