Proceedings of the Annual Meeting of the Cognitive Science Society

Recent work with brain-damaged pationts has provided evidence for a tentative neuroanatomical localization of mental image generatino in the posterior left hemisphere. This evidence will be briefly summarized and critiqued. And a new test of the localization, using normal subjects, will be presented. When mental images of stimuli were used a templates to facilitate a visual discrimination, the effect of imagery was greater for stimuli presented in the right visual field (left hemisphere) than in the left visual field (right hemisphere). This result is discussed in relation to earlier claims about the hemisphericity of imagery.

Conncetionist learning schemes have hitherto seemed far removed from cognitive skills such as reasoning, planning, and the formation of internal models. In this article we investigate what sort of world models a conncetionist system might learn and how it might do so. A learning scheme is presented that forms such models based on obserrved stimulus-stimuluus relationships. The basis of the scheme is a recurrently connected network of simple, neuron-like processing elements. The net produces a set of predictions of future stimuli based on the current stimuli, where these predictions are based on a model andn involve multiple-step chains of predictions . Results are presented from computer simulatinos of the scheme connected ot a simple world consisting of a stochasitc maze (Markov process). By wandering around the maze the network learns its construction. When reinforcement is subsequently introduced, the solution to the maze is learned much more quickly than it is without the exploration period. The form and logic of the experiment is the same as that of the latent learning experiments of animal learning research

Determining which features in an environment are salient given a task, salience assignment, is a central problem in machine learning. A related phenomenon, contingency ( the conditions under which relative salience among environemental features is acquired), is central to learning and memory in animal psychology. This paper presents an analysis of a set of empirical data on contingency and an algorithm for the salience assignment problem. The algorithm presented is implmented in a working computer profram which interacts with a simulated environement to produce contingent asssociative learning corresponding to relevant behavioral data. The model also makes specific empirical predictions that can be experimentally tested.

We argue that causal reasoning is an essential part of intelligent human behavior, and that discussion of it cannot be divorced from discussion of temporal reasoning. We therefore set out to define causation in three stages. In the first, we present an ontology of time. We then outline a theory of "causal conditional", which allows one to reason about multiple possible courses of events. Finally, we define causation in terms of direct e=causation and causal origins

The generation of extended plots for melodramatic fiction is an interesting Artificial Intelligence task --- one that requires the application of generalization techniques to carry out fully, UNIVERSE is a story-telling program that uses plan-like units, "plot fragments" to generate plot outlines. By using a rich library of plot fragments and a well-developed set of characters, UNIVERSE can create a wide range of plot outlines. In this paperm we illustrate how UNIVERSE's plot fragment library might be automatically extended using explanation-based generalization methods. Our methods are based on analysis of a televeision melodrama

This paper examines the phenomenon of daydreaming:spontaneously recalling or imagining personal or vicarious experiences in the past or future. The following important roles of daydreaming in human cognition are postulated; plan preparation and rehearsal, learning from failures and successes, support for processes of creativity, emotion regulation and motivation. A computational theory of daydreaming and its implementation as the program DAYDREAMER are presented. DAYDREAMER consists of 1) a scenario generator based on relaxed a planning, 2) a dynamic episodic memory of experiences used by the scenario generator based on relaxed planning, 2) a dynamic episodic memory of experiences used by the scenario generator based on relaxed planning, 2) a dynamic episodic memory of experiences used by the scenario generator, 3) a collection of personal goals and control goals which guide the scenario generator, 4) an emotion component in which daydreams initiate, and are initiated by emotional states arising from goal outcomes, and 5) domain knowledge of interpersonal relations and common everyday occurences. The role of emotions and control goals in daydreaming is discussed. Four control goals commonly used in guiding daydreaming are presented: rationaliszation, failure/success reversal, revenge, and preparation. The role of episodic memory in daydreaming is considered, including how daydreamed information is incorporated into memory and leter sed An initial version of DAYDREAMER which produces several daydreams ()

Recent artificial intelligence models of analogical reasoning are based on mapping some underlying causal network of relations between analogous situations. However, causal realtions relevant for the purpose of one analogy may be irrelevant for another. We describe here a technique wich uses an explicit representation of the putpose of the analogy to automatically create the relevant causal network. We illustrate the technique with two case studies in which concepts of everyday artifacts are learned by analogy

We discuss learning and the adaptive generation of concrete strategies through interactive experience. The domain is the game Tictactoe. The knowledge structures embodying strategies we represent as having tree parts: a Goal, a sequence of Actions, and a set of Constraints on those actions (GAC). We simulate such structures in a program that plays Tictactoe against different kinds of opponents. Applying these strategies leads to moves that often result in winning or losing; which in turn leads to the creation of new structures, by modifying the current GACs. These modifications are controlled by a small set of specific rules, so that the GACs are related by the ways modifications can map from one to another. Subject to certain limitations, we do a complete exploration of certain classes of strategy. This learnability analysis takes guidance from previous cognitive studies of a human subject by Lawler. The simulations were performed on a Symbolics 3600 in LISP. This work avoids abstractions in order to explore learning

We have proposed a neural network style model of language processing in a na effort to build a cognitive model which would simultaneously satisfy constraints from psychology and neurophysiology. This model was successful in disambiguating word senses in semantically determined sentences, but was unable to distinguish Agent from Object in semantically reversible sentences such as "John loves Mary." In this paper we rectify the matter by specifying the syntactic portion of the model, which is a massively parallel, completely distributed connectionist parser. We also describe the results of a simulation of the model.

We describe a connectionist parsing scheme based on context-free grammar rules. In this scheme we use an updating rule similar to the one used in the Boltzmann machine (Fahlman, Hinton and Sejnowski 1983) and apply simulated annealing. We show that at low temperatures the time average of the visited states at thermal equilibrium represents thte correct parse of the input sentence. In contrast with previously proposed connectionist schemes for natural alnguage processing, this scheme handles the traditionally sequential rule-based parsing in a general manner in the network. Another difference is the use of the computational scheme of the Boltzmann machine. This allows us to formulate general rules for the setting of weights and thresholds in our system. The parsing scheme is built from a small set of connectionist primitives that represent the grammar rules. These primitives are linked together using pairs of computing units that behave like discrete switches. These units are used as binders between concepts represented in collection of rules, and are very useful in the construction of connectionist schemes for any form of rule-based processing.

In this paper , we present a process model of editorial comprehension, representation, and retrieval. Editorial comprehension involves building an argument graph organized by abstract argument strategies which are represented declaratively as argument units. Issues include: (a) organizing and indexing goals, plans, events, states, beliefs, and belief justifications instantiated during comprehension of editorial arguments; and (b) retrieving information from conceptual representations of editorial arguments. This process model of reasoning and argument comprehension is currently being implemented in OpEd (Alvarado et al., 1985a), a computer system that reads short politico-economic editorials and answers questions about them.

Analogical resasoning is an important part of human intelligence. We often employ it as a vehivle for conveying ideas, and we rely upon it whenever we make a decision about a new situation [STER77]. This paper presents a theory of analogy recognition and comprehension, using as a domain letters to the editors of weekly news magazines. Our theory relies on lexical clues and the comparison of conceptual similarities to trigger recognition of the analogies in these letters. Our conceptual representation of an analogy in memory utilises comparison links to map analogous elements to each other and to tie together parallel arguments. We demonstrate application of this theory to a prototypical letter. The current status of a program implementing this theory is reviewed, and future research directions are discussed.

Two approaches to reasoning about uncertainty are discussed. A parallel certainty inference model superimposes reasoning about the credibility of inferences on a deductive framework. A second approach identifies and implements representativeness as the general determinant of credibility in classification tasks. These approaches are steps in the evolution of endorsement-based reasoning, a view of uncertainty in terms of structured objects that represent characteristics of evidence

This paper describes a proposal and some preliminary evidence in support of a model for understanding the points of simple stories. The proposed model differs from existing systems in that it includes, in addition to a representation of the plans and goals of each of the story characters, a model of the beliefs and intentions of the author and the reader. It is hypothesized that readers use story-specific information in conjunction with their own beliefs about the story events in order to make inferences relevant to the point of the story that the author intended. Evidence from adult readers is presented in support of each of the components of the model and their interaction. The proposed model has relevance for psychological and computational research on story understanding. The work also has implications for more general discourse situations in which understanding is predicated on the knowledge of shared beliefs.

Reasoning about changes of existence in objects, such as steam appearing and water disappearing when boiling occurs, is something people do every day. Discovering methods to reason about such changes in existence is a central problem in Native Physiscs. This paper analyzes the problem by isolating an important case, called quantity-conditioned existence. and presnets a general method for solving it. An example generated by an implemented program is exhibited, and remaining open problems are discussed

This short report summarizes a new model of question answering that we have developed and tested. The model specifies how humans answer many different kinds of questions (including why, how, when, where, enablement, consequence, and significane questions) after comprehending narrative passages. For example, if a narrative passage contained the episode the dragon kidnapped the maidens, te why-question for this episode would be why did the dragon kidnap the maidens? and possible answers would be because the dragon wanted to eat the maidens and because the dragon was lonely. According to the model, the major information sources for answers to questions include the passage structure and the generic knowledge structures that are associated with the content words in the query (i.e., DRAGON, MAIDEN, KIDNAPPING). After these knowledge structures are activated in working memory, there are convergence mechanisms which narrow down the node space to a set of relevant answers to a given question. The convergence mechanisms involve four major components. First, there is an arc search procedure, associated with each question category, which specifies what categories and paths of arcs are sampled when knowledge structures are tapped for answers. Second, there are a set of heuristics for establishing priorities among knowledge structures. heuristics for establishing priorities among knowledge structures. Third, there is an intersecting node identifier which segregates those nodes in a given knowledge structure which overlap (match) a node in at least one other knowledge structure in working memory. Fourth there is a constraint proagation component which prunes out erroneous nodes during the evaluation of (a) the intersecting nodes and (b) the nodes that radiate from intersecting nodes. The model has been tested by simulating question answering protocols collected from human subjects.

Anaphors, such as definite noun phrases and pronouns, are important contributors to discourse coherence. Anaphora resolution is the process of determining the referent of an anaphor in a discourse or dialogue. Models of discourse and sentence comprehension have made different claims about the temporal relationship between the occurence of the syntactic and semantic analyses of the sentence and the process of anaphora resolution. The end-of-sentence hypothesis holds that anaphora resolution occurs at the end of the sentence, after the syntactic and semantic analyses are completed. The immediacy assumption holds that anaphora resolution occurs as soon.as an anaphor is encountered and is completed as m u c h as possible before further words are processed. The cognitive lag hypothesis assumes that anaphora resolution starts when the anaphor is encountered but is completed while processing further words in the sentence. A study is described that traces the activation of a referent by its anaphor over a complete sentence. It demonstrates that anaphora resolution does not await the complete syntactic and semantic interpretations of the sentence. A n anaphor starts activating its referent as soon as the anaphor is encountered and the referent stays activated until the end of the sentence. This result supports a particular version of the immediacy assumption. This is also interpreted in terms of a limited cache that stores the items currently in focus and that is updated at sentence or clause boundaries.

A general purpose model and learning program are described which account for the phenomena of

A model of memory and learning is presented which indexes a new event by those features which are relevant in explaining why the event occurred. As events are added to memory, generalizations are created which describe and explain similiarities and differences between events. The memory is organized so that when an event is added, events with similar features are noticed. An explanation process attempts to explain the similar features. If an explanation is found, a generalized event is created to organize the similar events and the explanation is stored with the generalized event.

The paper reports recent results from the theory of Bayesian networks, which offer a viable formalizsm for realizing the computational objectives of connectionist models of knowledge. In particular, we show that the Bayseian network formalism is supportive of self-activated, multidirectional proagation of evidence that converges rapidly to a globally-consistent equilibrium.

We will first agure that ignoring possible deception in multi-agent scenarios can lead to planning failures; specifically, we show how standard deduction may be able to solve the Wise Man Problem, but not a variant where some agents are deceptive (i.e., the Wise-Yet-Deceitful Man Problem, or W-Y-D). Second we will show how to avoid planning failures in scenarios such as W-Y_D, by developing models of both(1) the deceptive tendedncies of other agents, and (2) how these other agents themselves reason about deception; the concepts of best-case and worst-case deceptive agents witll be introduced as examples. Third, we will

A computational model of learning in a complex domain is described an its implementation is discussed. The model supports knowledge-based acquisition of problem-solving concepts from observed examples, in the domain of physics problem-solving. The system currently learns aobut momentum conservation, in a psychologically plausible fashion form a background knowledge of Newton's laws and the calculus. In its contribution to machine learning, this research is important for artifical intelligence. From a psychological perspective it demonstrates the computational consistency of a machanism tha tmay underlie human learning iin a complex domain. This work also has implications for computer-adied instruction, in that it advances a learning model for a complicated domain involving both symbolic and numerical reasoning.

In this paper we present a process model that uses past experience in generating arguments of persuasion. We view persuasive argumentation as an instance of problem solving. As such, we employ knowledge organization idesas and problem solving techniques that have been advocated in an analogical view of problem solving. To illustrate our ideas, we use the domain of mediation of labor disputes. Our model is implemented in the PERSUADER, a computer program that gives advice in collective bargaining meditation.

This paper examines the process of storytelling and story invention. It focuses on the use of themes, episodic memory, analogical mappings, planning and literary goals. A computational model of storytelling is presented and its implentation as the program MINSTREL is discussed. MINSTREL contains the episodic memory of stories and themes an duses these memories along with the knowledge about the world of King Arthur's knights to invent interesting new stories.

Theories of discourse comprehension and memory for text usually assume a propositional format in which information is stored. In agreement with the work on mentla imagery we argue that information from textes may also be remembered in a spatial representation. Inference processes with spatial relations depend on the format of the mental representation. In two experiments we employed a priming technique to show spatial properties of mental representation. The first one using narratives failed to yield positive results. Th esecond experiment using spatial descriptions supported the hypothesis. Decision times in a priming task were dependent on spatial distances. The relationship between inference processes and the form of the mental representation is discussed.

Simple and complex addition problems were presented for true/false verification to 22 subjects across two times of measurement to test the general model for simple and complex addition proposed by Widaman, Cormier, & Geary (1985). Models fit to average RT data revealed that subjects were processing complex problems columnwise, beginning with the units column. Column sums seemed to be obtained through an incrementing process, and subjects exited problems as soon as a colimn error was encountered. Group level models were the same across complex problem types and for both times of measurement. However, individual level analyses suggested that nearly half of the subjects used a different processing strategy to obtain column sums for the second time of measurement. Results support the multi-staged model proposed by Widaman et al. (1985), but individual level results suggest that information processing models developed from group data may not represent the processing strategies used by all subjects, or the same subjects at different times.