Elizabeth Downs’s research while affiliated with Soar Technology, Inc. and other places

In many contexts, people generate forecasts about events of interest, and decision-makers wish to aggregate these forecasts to improve their accuracy. These forecasts differ from signals in the physical sciences. In particular, sensor signals are noisy samples from a common underlying distribution, while human-generated forecasts are based on cognitive models that vary from one informant to another. As a result, human forecasts, unlike physical signals, are not guaranteed to be statisticallyindependent conditioned on the true outcome. These differences both provide new opportunities for aggregation, and impose restrictions that do not apply to physical signals. This paper describes the difference between forecasts and physical signals, outlines a strategy for exploiting these differences in aggregation, and demonstrates modest but statistically significant gains in the accuracy of aggregated forecasts using data from a large ongoing experiment in forecasting world events.

Researchers have explored many formalisms to model how people think about their world. We describe an application that requires modeling how people forecast events in the real world. The naïve assumption is that they use formalisms that model how the world actually evolves. These formalisms are at variance with empirical psychological results. We present a more realistic alternative, the Narrative Space Model (NSM), describe a swarming agent algorithm to fit its parameters from observed data, and present some early results.

There is strong theoretical evidence that aggregation of human judgments should not simply average multiple forecasts together (the unweighted linear opinion pool, or ULinOP), but weight them in such a way as to insure representation of a maximally diverse set of models among the experts from whom they are elicited. Explicitly eliciting these models places a major burden on me experts. We report on a variety of approaches to estimating these models, or at least the diversity among them, with minimal explicit input from the experts. Copyright © 2012, Association for the Advancement of Artificial Intelligence (www.aaai.org). All rights reserved.

A number of studies have explored the dynamics of opinion change among interacting knowledge workers, using different modeling techniques. We are particularly interested in the transition from cognitive convergence (a positive group phenomenon) to collapse (which can lead to overlooking critical information). This paper extends previous agent-based studies of this subject in two directions. First, we allow agents to belong to distinct social groups and explore the effect of varying degrees of within-group affinity. Second, we provide exogenous drivers of agent opinion in the form of a dynamic set of documents that they may query. We exhibit a metastable configuration of this system with three distinct phases, and develop an operational metric for distinguishing convergence from collapse in the final phase. Then we use this metric to explore the system's dynamics, over the space defined by social affinity and precision of queries against documents, and under a range of different functions for the influence that an interaction partner has on an agent.

At a tactical level, insurgents planning attacks with Improvised Explosive Devices (IEDs) are constrained in their choice of target by the specific location of their safe house or weapons cache, the geographic context in which they operate, and the pattern of potential targets as it presents itself at a given time. Geographic profiling in law-enforcement already takes advantage of similar constraints to identify possible origin locations of serial offenders. We show how geographic profiling of past IED events can significantly enhance our ability to identify areas at risk for future attacks. Specifically, we introduce three tightly coupled swarming pattern analysis models (profiling, clustering, forecasting) that refine each others' conclusions dynamically and point to systematic evaluation experiments that confirm the research hypothesis.

Swarming agents often operate in benign geographic topologies that let them explore alternative trajectories with minor variations that the agent dynamics then amplify for improved performance. In this paper we demonstrate the deployment of swarming agents in the non-metric and discontinuous topology of a process graph. We align our research with traditional AI approaches and focus on hierarchical task network (HTN) descriptions of constraints and preferences in the execution of abstract methods by a group of real-world entities. In particular, we adapt the TAEMS representation to place a greater emphasis on the mediation of method-execution through shared resources and collectively achieved quality (stigmergic coordination). The paper presents our polyagent model and experiments that demonstrate the scalability of the system and the ability of our agents to achieve optimal entity coordination.

Stigmergy is usually used to model semantically simple problems such as routing. It can be applied to more complex problems by encoding them in the stigmergic environment. We demonstrate this approach by showing how stigmergic agents can plan over a hierarchical task network, specifically a resource-oriented dialect of the TÆMS language.

Stigmergy, usually used on simple problems, can be applied to more complex ones by encoding them in the agents' environment. We show how stigmergic agents can plan over a hierarchical task network, a resource-oriented dialect of the TÆMS language. Our results reveal an important distinction among HTN's.

Self-organizing systems are robust, scalable, adaptive to a changing environment, and tolerant to noise and incomplete or conflicting information. These are the requirements for our information matching system (IMS) that organizes models of document contents and user interest in an abstract information space by relevance to provide any-time recommendations of other users (for collaboration) or documents (for information gathering) to intelligence analysts. In this report on research-in-progress, we present a plug-and-play integration architecture for multiple and possibly competing modelers of arbitrary (text, audio, video, etc) document contents that influence the emerging arrangement of document and user models. The contributions of these modelers are numerical similarity statements that specify attractive or repulsive forces, which guide the ongoing rearrangement of the current set of models. This self-organizing force-based arrangement process adjusts dynamically to changes in the document set or shifting user interest. Our paper also discusses related research, initial experiments that indicate satisfactory system-level behavior, and an upcoming evaluation exercise with actual users.

A number of studies have explored the dynamics of opinion change among interact-ing knowledge workers. We are particularly interested in the transition from cognitive convergence (a positive group phenomenon) to collapse (which can lead to overlooking critical information). This paper extends previous agent-based studies of this subject in two directions. First, we allow agents to belong to distinct social groups and explore the effect of varying degrees of within-group affinity. Second, we provide exogenous drivers of agent opinion in the form of a dynamic set of documents that they may query. We exhibit a metastable configuration of this system with three distinct phases, and develop an operational metric for distinguishing convergence from collapse in the final phase. Then we use this metric to explore the system's dynamics, over the space defined by social affinity and precision of queries against documents, and under a range of different functions for the influence that an interaction partner has on an agent.