Soft information fusion, fusing information from natural language messages with other soft information and with information from physical sensors is facilitated by representing the information in the messages as a formally defined propositional graph that abides by the uniqueness principle- the principle that every entity or event that is mentioned in the message is represented by a unique node in the graph, or, at worst, by several nodes connected by co- referentiality relations. To further facilitate information fusion, information from the message is enhanced with relevant information from background knowledge sources. What knowledge is relevant is determined by also representing the background knowledge as a propositional graph, embedding the knowledge graph from the messages into the background knowledge graph using the uniqueness principle to fuse a message graph node with a background knowledge graph node, and then using spreading activation to find subgraphs of the background knowledge graph. This combination of the message graph with the retrieved subgraphs is considered the " relevant information." In this chapter, we discuss, evaluate, and compare two techniques for spreading activation.

A soft-information fusion process produces refined estimates of soft-information, such as natural language messages. Information resulting from a soft-information process can be used to retrieve related, relevant information from background (a-priori) knowledge sources using contextual “cues” contained in those messages, a process we call “Context-Based Information Retrieval (CBIR)”. These retrieval results can be used to aid further understanding, and other fusion operations (e.g., data association). CBIR process performance is dependent on the choice of algorithms and parameters for those algorithms, and it is crucial that these are chosen appropriately for the problem domain the CBIR algorithm is used to aid. In this paper an f-measure evaluation of two spreading activation algorithms and their parameters is given using a soft information fusion process in a counterinsurgency domain. This evaluation takes place in two phases. The first phase executes the algorithms over a range of values in order to determine how those parameters affect the performance of the algorithms, and to set these parameters for future use. The second phase compares the results of these algorithms using the parameter settings learned from the first phase.

This paper presents a soft information fusion framework for creating a propositional graph from natural language messages with an emphasis on producing these graphs for fusion with other messages. The framework utilizes artificial intelligence techniques from natural language understanding, knowledge representation, and information retrieval.

Contextual Information is proving to be not only an additional exploitable information source for improving entity and situational estimates in certain Information Fusion systems, but can also be the entire focus of estimation for such systems as those directed to Ambient Intelligence (AI) and Context-Aware(CA) applications. This paper will discuss the role(s) of Contextual Information (CI) in a wide variety of IF applications to include AI, CA, Defense, and Cyber-security among possible others, the issues involved in designing strategies and techniques for CI use and exploitation, provide some exemplars of evolving CI use/exploitation designs on our current projects, and describe some general frameworks that are evolving in various application domains where CI is proving critical.

Computationally expensive processes, such as deductive reasoners, can suffer performance issues when they operate over large-scale data sets. The optimal procedure would allow reasoners to only operate on that information that is relevant. Procedures that approach such an ideal are necessary to accomplish the goal of commonsense reasoning, which is to endow an agent with enough background knowledge to behave intelligently. Despite the presence of some procedures for accomplishing this task one question remains unanswered: How does one measure the performance of procedures that bring relevant information to bear in KR systems? This paper answers this question by introducing two methods for measuring the performance of context-based information retrieval processes in the domain of KR systems. Both methods produce an f-measure as a result. These methods are evaluated with examples and discussion in order to determine which is more effective. Uses of these measures are also discussed.

We present a categorization of contextual constraints, and discuss their uses in embodied agent architectures. \Context" has been described as a dicult term to dene, because it’s: (1) used across numerous disciplines in cognitive science and computer science; (2) relative to an agent, or device; and (3) relative to the cognitive process being examined and experimented upon. As such, context is a consequence of theories about cognitive processes, not something observed. It has a theoretical role, not one of a measurable unit. We will take context to be the structured set of variable, external constraints to some (natural or articial ) cognitive process that inuenc es the behavior of that process in the agent(s) under consideration. By reviewing the cognitive science disciplines of linguistics, psychology, knowledge representation, and human-computer interaction, we’ve identied contextual factors that can serve several uses among embodied cognitive architectures, such as knowledge acquisition, knowledge partitioning, and context switching.

Cyber Security can benefit greatly from the association and combination of data and information from multiple sources. A data repository of system vulnerabilities, a network scanning tool, and the advice of a systems analyst trained in cyber security can all aid in identifying and preventing intruders. Previous attempts at information fusion in cyber security have largely concerned themselves with the tangible information sources, but this ignores an important resource in solving problems in this particular domain --- the cyber security expert's reasoning process. The National Center for Information Fusion (NCMIF) has begun implementing a solution that partially automates the cyber security expert in the intrusion detection process through a combination of information fusion techniques and symbolic reasoning, using the SNePS knowledge representation, reasoning, and acting system. Our methodology approaches cyber security problems by fusing information from external information repositories into a SNePS-based agent‟s knowledge base. We have identified five information sources that are useful: the background knowledge of a cyber security subject matter expert (SME); Nessus security scan reports; the Common Vulnerabilities and Exposures (CVE) database; and INFERD template graphs. The SNePS system makes use of higher-order logic to represent information about the external world. Facts are represented as proposition-valued terms, and the SME‟s reasoning procedures are represented as logical rules.

The SNePS knowledge representation, reasoning, and acting system has several features that facilitate metacognition in SNePS-based agents. The most prominent is the fact that propositions are represented in SNePS as terms rather than as sentences, so that propositions can occur as arguments of propositions and other expressions without leaving first-order logic. The SNePS acting subsystem is integrated with the SNePS reasoning subsystem in such a way that: there are acts that affect what an agent believes; there are acts that specify knowledge-contingent acts and lack-of-knowledge acts; there are policies that serve as "daemons," triggering acts when certain propositions are believed or wondered about. The GLAIR agent architecture supports metacognition by specifying a location for the source of self-awareness and of a sense of situatedness in the world. Several SNePS-based agents have taken advantage of these facilities to engage in self-awareness and metacognition.

We present a solution to McCarthy’s Second Telephone Number Problem. This problem requires an agent to: realize that it lacks some knowledge to complete a task; know the external knowledge sources it can use to obtain the knowledge; know how to obtain the missing knowledge from those sources; actually obtain the missing knowledge; and use the obtained knowledge to complete the task. We explain how a SNePS-based agent solves the problem, and the features of SNePS and the GLAIR architecture that facilitate the solution. These features include: the integration of reasoning and acting; the ability to reason about actions; and the ability to represent and reason about the knowledge contained in external knowledge sources. This latter feature is accomplished through the process of grounding of symbols of the knowledge representation in symbols of an external language.

In this paper we sketch a new approach for agent design and operability called co-designing agents (CDA). As a working definition, we take a CDA to be an agent that participates in some aspect of its own design. The precise manner and degree of a CDA’s participation will ultimately be a design-time decision, but by considering a specific agent implementation (that of AI actor-agents in a virtual drama) we are able to extract a general set of CDA requirements. The CDA approach utilizes concepts from several areas of active research in AI. We present a broad summary of the relevant literature and discuss its applicability to CDA design. We then consider the SNePS knowledge representation, reasoning, and acting system as a potential CDA implementation platform.

We demonstrate the use of SNeRE, the acting component of the SNePS knowledge representation, reasoning, and acting system, by showing its use to implement a wumpus world agent [Russell and Norvig, 1995],. For this purpose, we use SNePS 2.6.2, which consists of SNePS 2.6.1 [Shapiro et al., 2004] plus some patch files. We usually name our SNePS-based agents Cassie [Shapiro, 1989; 1998; Shapiro and Is-mail, 2003; Shapiro et al., 2000; Shapiro and Rapaport, 1987; 1991]. To distinguish Cassie in the role of the wumpus world agent, we will call her CassieW. Our main motivation in developing intelligent systems is to model general human-level intelligence, not to maximize the use of computing power to optimize problem solving. CassieW has been developed accordingly.

We demonstrate the use of SNeRE, the acting component of the SNePS knowledge representation, reasoning, and acting system, by showing its use to implement a wumpus world agent [Russell and Norvig, 1995]1. For this purpose, we use SNePS 2.6.2, which consists of SNePS 2.6.1 [Shapiro et al., 2004] plus some patch files. We usually name our SNePSbased agents Cassie [Shapiro, 1989; 1998; Shapiro and Ismail, 2003; Shapiro et al., 2000; Shapiro and Rapaport, 1987; 1991]. To distinguish Cassie in the role of the wumpus world agent, we will call her CassieW. Our main motivation in developing intelligent systems is to model general human-level intelligence, not to maximize the use of computing power to optimize problem solving. CassieW has been developed accordingly.

We are demonstrating several intelligent agents built according to the MGLAIR (Modal Grounded Layered Architecture with Integrated Reasoning) agent architecture. The top layer of MGLAIR is implemented in SNePS and its acting subsystem, SNeRE (the SNePS Rational Engine). The major demonstration will be act 3 of The Trial The Trail, an interactive drama running on an immersive Virtual Reality system, in which a human participant interacts with several MGLAIR actor-agents. We will also demonstrate several olher MGLAIR agents that operate in non-VR graphical environments. All these agents illustrate our approach to building agents with integrated first-person, on-line reasoning and acting.

The Trail is an interactive drama running on an immersive VR system. Imagine Tarkovsky's Stalker, crossed with Alice Through the Looking Glass, crossed with Monty Python and the Holy Grail. Now imagine embarking on a guided journey through this warped yet familiar landscape. Your guides are two intelligent agents, Patofil and Filopat. We consider a virtual reality drama to be a scripted play in which the computational agents are actors who have copies of the script, and one human audience member has been drafted to be a participant, but doesn't have a copy of the script. The computational actors must improvise reactions to the human participant's actions, but keep the play moving along in as close agreement to the script as possible. The goal is to provide the human participant with a specific emotional experience.

We provide an overview of the use of intelligent agents, implemented in the new MGLAIR architecture, in a virtual reality drama. For us, a virtual reality drama is a scripted play in which the computational agents are actors who have copies of the script, and one human audience member has been drafted to be a participant, but doesn’t have a copy of the script. The computational actors must improvise reactions to the human partpicipant’s actions, but keep the play moving along in as close agreement to the script as possible. The goal is to provide the human participant with a specific emotional experience. We explicate this philosophy; outline the previously described GLAIR architecture; explain the introduction of an organization into modalities that results in the new MGLAIR architecture; describe our current VR drama, The Trial, The Trail; and discuss the implementation of our actor-agents. Our discussion of the actor-agents focuses on their abilities to react to triggers (cues), their performance of contingent actions that are off the main-line arc of the script, their use of timers to pace the drama, and the organization of the cast of actor-agents into a multi-agent system.

This paper describes the implementation of the Shakey robot on a Magellan Pro hardware robot. The paper gives an account of the SNePS/GLAIR architecture which has been implemented for the Magellan Pro. It details a challenging task of reinventing Shakey in order to expand the horizons he has inhabited so far. The aim is to utilize prior knowledge about SNePS, build a GLAIR architecture and integrate this with the capabilities of the Magellan Pro so that we have a robot capable of sensing, planning and acting intelligently. The paper also describes in detail the Knowledge Level of the GLAIR architecture which is implemented in SNePS for the Magellan Pro hardware robot.

A simple vacuum-cleaner agent is introduced in Russell and Norvig’s artificial intelligence (AI) text [Russell and Norvig, 2003] to illustrate different agent types to beginning AI students. Underlying the different agent types are several simple actions that take place in a vacuum cleaner world consisting of a grid of squares, some of which contain dirt. The agent’s actions include turning left or right, moving forward, and picking up dirt. Having students write a program to simulate the vacuum world is a useful way to provide them with a feeling for different agent types in a simplified environment. However, implementing a vacuum-cleaner agent using a low-cost robotics kit might teach students much more about agents in the real world and could serve to get them interested in and excited about AI in a way that working with purely simulated environments may not. This paper describes the design and implementation of “Dustbot”, a robot based on the Russell and Norvig vacuum-cleaner agent. The Dustbot project was carried out as part of an independent study by a student who had already taken the undergraduate AI course. The purpose of the project was to test and debug the vacuumcleaner robot and to develop a set of instructions that could be used in subsequent offerings of our undergraduate AI course.