Frederick Hayes-Roth

Network Structures for Distributed Situation Assessment

A new approach to situation assessment is an automated distributed sensor network (DSN) consisting of many intelligent sensor devices that can pool their knowledge to achieve an accurate overall assessment of a situation. Laboratory experiments were conducted to investigate potential DSN organizations and to ascertain some general design principles. These experiments have been performed with a network of sensor nodes, each of whom sees only a small portion of the entire environment and attempts to identify the environmental mobile entities as quickly as possible. To do this, they must cooperatively communicate their hypotheses and data, using a limited number of messages. Two general DSN organizations were tested. The first was hierarchical. The second was an anarchic committee whose nodes could each send messages to one, some, or all other nodes. The performance of the committee organization consistently surpassed the hierarchical one. This lent support to the contention that DSN architectures need to emphasize cooperative aspects of problem-solving. A machine-based simulation of such a network that achieved performance levels comparable to that of the human committee DSN organization was also constructed and tested. Because most situation assessment communications concern hypothesis updating and revision, minimizing communication requirements through the concept of active hypothesis processes, which are responsible for predicting their own evolution over time, is suggested.

Plasticity in Memorial Networks.

An adaptive network model is proposed to represent the structure and processing of knowledge. Subjects formed artificial knowledge structures through controlled storage and retrieval events. The accessibility of stored information was measured as reaction time for verifying relations between memorized concepts, before and after systematic manipulation of subjects retrieval experiences. The results indicate relationships among (a) the frequency of verifying a test relation, (b) the number of other relations involving the concepts used to evaluate the test relation, (c) the recency and frequency of verifying those other relations, and (d) reaction time to verify the test relation. Subjects judgments of distance between pairs of memorized concepts were used to produce a theoretically reasonable spatial representation of the acquired knowledge structure.

Distributed Intelligence for Air Fleet Control

Abstract : Distributed planning and control techniques provide potential advantages over centralized porcessing in speed, cost, reliability, flexibility, and minimization of long-distance communications for a variety of tasks, including military threat assessment, command and control decisionmaking, disaster relief coordination, and civilian air traffic control. Six different architectures for distribution of control among multiple processors are presented, and the influences of different task environments on each are discussed. The work focuses primarily on the use of distributed planning and control for civilian air traffic control. One architecture, in which each aircraft is controlled by a separate processor, is described in detail in an illustrative scenario. An initial system design is presented in which cooperating experts comprise a processor. These cooperating experts sense or infer aircraft intentions, generate and evaluate plans, control and monitor plan execution, and communicate with other processors.

Focus of attention in a distributed-logic speech understanding system

The Hearsay II speech understanding system under development at Carnegie-Mellon University is a complex, distributed-logic processing system: Processing in the system is affected by independent, data-directed knowledge sources processes which examine and alter values in a global data base representing hypothesized phones, phonemes, syllables, words, and phrases, as well as the hypothetical temporal and logical relationships among them. The question of how to schedule the numerous potential activities of the knowledge sources so as to understand the utterance in minimal time is called the focus of attention problem. Near optimal focusing is especially important in a speech understanding system because of the very large solution space that potentially needs to be searched. Using the concepts of stimulus and response frames of scheduled knowledge source instantiations, competition among alternative responses, goals, and the desirability of a knowledge source instantiation, a general attentional control mechanism is developed. This general focusing mechanism facilitates the experimental evaluation of a variety of specific attentional control policies (such as best-first, bottom-up, and top-down search heuristics) and allows the modular addition of specialized heuristics for the speech understanding task.

Representation of structured events and efficient procedures for their recognition

Abstract Structured events are configurations of objects in logical, spatial, temporal or activity relations. A parameterized structural representation system for this class of events is discussed. Parameters in such representations are arbitrarily chosen symbols used to insure consistent references to the same object in diverse relations. All-or-none matching of two representations is the basis for pattern recognition. In this framework, descriptions of pattern or concept prototypes act as structural templates for stimuli. As a result, recognition can be performed in a natural and structural way and is unaffected by manipulations of irrelevant variables. Typical recognition procedures are reviewed and a variety of alternative approaches are considered in light of the potential combinatorial explosions which might arise in applications of these procedures. One alternative is proposed which can exploit both redundancy among partially matching templates and computational parallelism in exhaustive search (recognition) problems. Another possibility considered is to find special recognition procedures for particular recognition problems. For example, to accomplish word recognition in speech understanding systems, highly practical techniques exist to match many templates in parallel (simultaneously) using only simple bit string operations. In addition, both the possibility of additional heuristic approaches to general recognition procedures and the use of less general relational representations are also considered in this paper.

Schematic classification problems and their solution

Abstract The necessity arises in a variety of tasks to classify items on the basis of the presence of one of a number of criterial sets of co-related feature values. Such sets are called class characteristics. Because such classification problems require the identification of characteristics on the basis of limited training information, they entail a difficult search problem. Consideration of the differences between the theoretical models underlying characteristic and volume pattern generators suggests a schematic approach. Schemata, sets of commonly co-occuring features values, are probabilistic indicators of class membership whenever the characteristics are unknown but the characteristic model prevails. Formal and algorithmic solutions to the classification problem when exemplars are simple (consist only of M feature or attribute values) are described. The relevance of these procedures to problems involving general (relational) data structures is also indicated.

Selection of word islands in the Hearsay-II speech understanding system

In Hearsay-II, a word recognizer hypothesizes words bottom-up from acoustic data. Usually many competing words are hypothesized for each time interval of speech, with the correct word rarely top-ranked. Due to the unreliable ratings of words and the limited syntactic constraint supplied by single words, the use of single-word islands would cause the recognition system to explore many blind alleys before abandoning an incorrect island. In addition, the multiplicity of words makes the parsing of all possible word sequences extremely time-consuming. The Hearsay-II island selection strategy uses (1) knowledge of what word adjacencies are allowed by the grammar, (2) anaiysis of acoustic data at the junctures between word hypotheses, and (3) heuristics based on the number of competing word hypotheses, to form multi-word islands which the syntax-level knowledge source first checks for grammatically and then attempts to extend to form a complete recognition.

A production system for speech understanding

A production system was developed for syntactic and semantic processing within the Hearsay-II speech understanding system. The major properties of the system are discussed, including (1) conversion of static language descriptions into productions, (2) compilation and data-directed execution of productions, (3) dynamically modifiable thresholds on goodness of pattern matches, (4) partial matching, and (5) representation and integration of bottom-up, top-down and horizontal searches. Several weaknesses of the production system paradigm in this application appeared during the course of this research. These arose from (1) the arbitrariness of canonical decompositions of patterns into subpatterns, (2) insufficient use of contextually confirming evidence for individual hypotheses due to the narrowness (literality) of monitored conditions, and (3) difficulty in evaluating the varying import of the same generic action in different contexts. Thus, while the uniformity and lack of explicit organization of production systems are touted as their most desirable features, attendant difficulties of dynamically organizing and controlling coherent problem solutions must be seriously considered in problem domains requiring careful allocation of computational resources.

Inferential searches of knowledge networks as an approach to extensible language understanding systems

A program has been constructed that implements some inference schemes for inferring the meaning of noun-noun, adjective-noun, and agent-verb-object phrases from the constituent words. Dictionary definitions of words are input directly into a semantic network using a context free grammar implemented in a production system. To interpret a novel phrase (e.g., lawn mower, car tire, car wheel) semantically, an intersection search originating from the nodes representing the constituent words is performed in the semantic network. If an intersection is found, a meaning for the phrase is produced; otherwise the phrase is not interpreted. During this interpretation, only knowledge provided by dictionary definitions and parts of speech of words is used.

Syntax and semantics in a distributed speech understanding system

The Hearsay II speech understanding system being developed at Carnegie-Mellon University has an independent knowledge source module for each type of speech knowledge. Modules communicate by reading, writing, and modifying hypotheses about various constituents of the spoken utterance in a global data structure. The syntax and semantics module uses rules (productions) of four types: (1) recognition rules for generating a phrase hypothesis when its needed constituents have already been hypothesized; (2) prediction rules for inferring the likely presence of a word or phrase from previously recognized portions of the utterance; (3) respelling rules for hypothesizing the constituents of a predicted phrase; and (4) postdiction rules for supporting an existing hypothesis on the basis of additional confirming evidence. The rules are automatically generated from a declarative (Le., non-procedural) description of the grammar and semantics, and are embedded in a parallel recognition network for efficient retrieval of applicable rules. The current grammar uses a 450-word vocabulary and accepts simple English queries for an information retrieval system.