Jerome R. Busemeyer - US grants

We propose to investigate and model how humans learn intervening concepts. Intervening concepts, such as for example the psychological concept of hunger, seem to be an important and natural part of human cognition. In our paradigm, subjects will be exposed to multivariate input-output relations related by hidden intervening concepts. Subjects' global and trial-by-trial learning processes will be examined as subjects attempt to learn how to predict the outputs given a particular set of input values. In a series of nine experiments, we plan to determine (a) the conditions under which a hidden intervening concept can be learned solely on the basis of experience with the input-output relations, and (b) the mechanisms used to learn a hidden intervening concept. The proposed experiments are designed to obtain converging evidence from a wide range and variety of experimental manipulations and analytical methods. The experiments include manipulations of (a) training and transfer across different types of causal structures, (b) various instructional conditions, (b) various types of training sequences, (c) various types of stimulus conditions, and (d) the use of selection versus reception concept learning training procedures. Two new and different types of analyses will be used to analyze the results: (a) a model-free global (principle component) measure of intervening concept learning which will be used to examine the global effects of the experimental manipulations, and (b) micro-level analyses of the trial by trial learning process. One important feature of both methods of analyses is that they can be performed at the individual subject level, thus permitting an investigation of the implications of individual differences in learning of intervening concepts for successful transfer performance. Importantly, the results will be used to compare and contrast a passive adaptive network learning model versus an active hypothesis testing learning model.

DESCRIPTION (Adapted from the Applicant's Abstract): Most real life decisions involve multiple-stages, that is a sequence of actions and events over time. There is a well developed theory of decision making, called dynamic programming, that is used to prescribe the optimal way to make multiple-stage decisions. This theory is based on three fundamental principles: Dynamic consistency, consequentialism, and substitutability. However, most of the past experimental research investigating the fundamental principles of decision making has been based on single-stage decisions. Thus, the empirical validity of the three fundamental principles of dynamic programming remain untested, and there is a need for fundamental research on multiple-stage decisions. Dynamic programming theory is an optimal theory that is not based on psychological principles. An alternative psychological theory of decision making, called decision field theory, has important implications for multiple-stage decisions. Specifically, dynamic consistency and substitutability are predicted to be violated by human decision makers under certain conditions specified by the decision field theory. The purpose of the proposed research is to conduct three sets of experiments, where each set is designed to empirically test one of the three principles of dynamic programming. The experiments are designed to produce conditions predicted by decision field theory to lead to violations of two of the three principles.

The primary goal of this project is to explain long-term, complex change processes in human-bioecological systems-especially forested regions. We will develop agent-based models to examine how land-use decisions made at one level (a household) affect outcomes at that level and at several higher and lower levels in a hierarchically nested set of systems. We develop two agent-based models to explain land-use patterns in the frontier and post-frontier Midwest of the United States and the frontier of the Brazilian Amazon. The first model will address two major puzzles: (1) Why did the descendants of the initial settlers in nineteenth-century Indiana cut down timber at such a massive and seemingly uneconomic rate that they eventually denuded the land, causing massive erosion and soil loss, and leading to substantial farm abandonment? and (2) Why have forests regrown so extensively on privately owned land when so many public policies are based on the assumption that fragmented, privately owned parcels are destined never to have significant forest regrowth? The second model will explain the spatial and temporal patterns of deforestation in the Amazon over the last three decades. The assumptions we make in the two models will be empirically tested and grounded by rigorous laboratory experiments. The patterns of land use at any point in time and the processes of change also will be tested against a rich set of data derived from ground-truthed satellite data, aerial photographs, land surveys, census data, household interviews, forest mensuration undertaken in a sample of forest patches, and archival data regarding timber and agricultural prices, input costs, and land values. After further development and testing, both models will be used to extrapolate into the future and assess how diverse public policies are likely to affect land use in general and forest change in particular in these regions. The project will involve three important capstone activities: a Workshop on Agent-Based Models of Biocomplexity, a synthesis volume to be derived from the Workshop, and a Summer Institute.

The study will have multiple impacts. By achieving an empirically validated understanding of land-use decisions of individual households under different policy regimes, the study will produce useful tools for evaluating alternative public policies. Ascertaining how public inducements, taxation, and constraints affect rates of forest change contributes to the worldwide effort to find effective methods for stimulating reforestation and thereby sequestering carbon to offset carbon released into the atmosphere. The study also addresses fundamental questions related to the appropriate model of human behavior to use when examining a combination of investment decisions in complex, dynamic environments. Thus, the study is relevant for achieving an empirically validated foundation for an array of decision situations beyond those of land use and deforestation. Tools from multiple social, biological, and physical science disciplines will be combined and expanded in unique ways and disseminated in publications, workshops, and training institutes. This research activity was funded as part of the FY2000 Biocomplexity Special Competition.

This proposal aims to develop a unique combination of state-of-the-art spatial, visualization, and experimental capabilities within a behavioral science research laboratory facility at Indiana University. The laboratory will be multi-purpose, with a focus on understanding complex systems at multiple temporal and geographic scales. It will contribute to both research and instruction. The laboratory will enable the development of new capabilities for spatially organized agent-based modeling, three-dimensional (3-D) visualization of social systems, multi-user 3-D virtual desktop worlds, as well as the implementation of behavioral experiments and GIS technologies in social science research. The research group proposing this set of development activities is multi-disciplinary with strong records of field and laboratory research in anthropology, economics, geography, information science, political science, and psychology. An important aspect of the proposed laboratory facility is to provide the infrastructure for fostering stronger linkages among the researchers involved in this proposal. It will create opportunities for enlarging the community of scholars and students at Indiana University interested in frontier social science that incorporates spatial attributes of decision environments important for understanding human and biological dimensions of social, political, and economic decision making. The laboratory will enable faculty and students to participate in the testing of current social science theories with computer-simulated, agent-based systems as well as the tools of laboratory decision-making experiments in controlled environments. Based on the research conducted in this laboratory, we hope to develop new theories of how complex individual interactions over space and time lead to emergent properties in complex social-ecological systems.

The proposed laboratory will facilitate the productivity and breadth of several ongoing and future activities: (1) a recently funded biocomplexity project integrating agent-based modeling, laboratory experiments, and GIS technologies into the study of complex land-use decisions and emergent land-use patterns; (2) research on networks of political communication related to the accessibility of political attitudes, orientations, and judgments; (3) investigation of resource allocation decision making under risk and uncertainty; (4) multi-agent experiments in the laboratory, in the field, and at multiple sites; (5) the incorporation of GIS technologies into social science research integrating key incentive and outcome effects; (6) creation of artificial 3-D environments for experimental research and policy applications; (7) paleoanthropological research on human evolution; (8) international relations; and (9) education and outreach.

[unreadable] DESCRIPTION (provided by applicant): Human concepts are complex, varied, and serve myriad purposes. One way concepts are used is to categorize people or things and infer properties from category membership. Historically, this view of concepts has dominated the theoretical and empirical literature in cognitive psychology. But this view is too restrictive and another important way concepts are used is to make predictions from strengths of causes to magnitudes of effects on the basis of continuous functional relationships. The general purpose of the proposed research is to provide the foundations for a more formal, systematic, and integrative approach to function learning that parallels the existing progress in category learning. More specifically, we aim to achieve the following three specific goals. First, we plan to rigorously test rule versus associative based models of function learning in a restricted domain that includes only single input - single output functions. Our second line of research addresses the possibility that the rule-abstraction and exemplar-based processes that are observed by individuals in function learning indicate a general learning orientation that produces characteristic learning outcomes for a host of higher-order cognitive tasks. In our third major focus, we examine the interrelations between learning and decision-making by manipulating the type of payoffs provided as feedback during function learning. This last line of research is designed to test the basic learning mechanisms underlying almost all connectionist-learning approaches. [unreadable] [unreadable]

[unreadable] DESCRIPTION (provided by applicant): This is an application for competitive continuation of our research program, which takes a highly integrative approach to understanding drug abuse by using formal decision models to examine the psychopathology of addictions. In our first project period, we used a single stage decision task (Iowa gambling task) to examine individual differences in how learning, motivational, and choice mechanisms interact in the decision making of drug abusers. In the second project period, we will move beyond the single stage decision approach, to develop and empirically test a computational (formal and mathematical) model for the multi-stage self-control decision problem. Multi-stage decision problems more closely resemble the complex decisions drug users face in which they must make a series of decisions that have both proximal outcomes (i.e., getting high, spending money, relief from withdrawal symptoms) as well as possible long term outcomes (i.e., addiction, social and occupational dysfunction). The model addresses three components: Devaluation of immediate reward and delayed punishment; 2) learning of rewards and punishments from trial by trial experience; and, 3) the critical planning process needed to achieve self control in multiple stage decision scenarios. The project has three specific aims. First, we will determine how regular drug users differ from comparison groups on basic evaluative characteristics, including risk aversion, loss aversion, and temporal discounting. Second, we will examine how these evaluative processes interact with learning to guide behavior over time. Third, we will refine and empirically test our comprehensive model for the multiple stage self-control decision problem in regular drug users. The results of our computational model will be examined in the broader context of drug use, demographic, psychosocial, and personality factors that have been linked to drug abuse. The theoretical model from this work will provide a deeper understanding of individual differences in decision processes of drug abusers, will provide a valuable perspective for the prevention and intervention strategies. [unreadable] [unreadable]

Research on human judgment and decision making has revealed a number of paradoxical findings that have resisted explanation under a common theoretical framework. These include violations of the sure thing axiom of decision making, interactions between inferences and decisions, violations of the reduction axiom of decision making, violations of the conjunctive and disjunctive axioms of probability theory, and order effects on judgments. In the past, separate and disconnected explanations have been proposed using variants of classic decision theory. This research proposes a unifying explanation for all of these paradoxical results based on a new quantum decision theory. Classic decision theory is based on classic probability theory. Probabilities are assigned to events defined as subsets of a universal set, which obey all the laws of Boolean algebra. Quantum decision theory is based on quantum probability theory. Probabilities are assigned to events defined as subspaces of a Hilbert space, which obey all the laws of Boolean algebra except the distributive axiom. Following from the distributive axiom, classic probability theory adheres to one of its most important theorems, the law of total probability. Because quantum logic does not have to obey the distributive law, quantum probabilities do not have to obey the law of total probability. Instead, quantum probability theory must obey another law called the doubly stochastic law, which the classic probability model does not obey. Hence, the two probability theories are fundamentally different and the critical question is which set of rules provides a better description of human behavior. The immediate goal of this research is to rigorously compare decision models built upon classical probability theory with those built from quantum probability theory. To rigorously compare quantum versus classical probability models of decision making, a series of experiments will be conducted. The experiments focus on tests of the law of total probability and tests of the law of double stochasticity, where the two classes of models make major and qualitatively different predictions. The research will accomplish three objectives: (1) develop a new quantum theory of human inference and decision making, (2) conduct new empirical tests of the fundamental laws of total probability and double stochasticity using human inference and decision behavior, and (3) rigorously compare and contrast classic versus quantum models of decision making with respect to the new empirical findings. A quantum or classic model will be preferred only if it provides a superior scientific explanation of the phenomena with respect to both accuracy and parsimony.

The broad and long-term goal of this research program is to break new ground and pioneer a new path by building probabilistic and dynamic systems for social and behavioral sciences from quantum rather than classical probability principles. Previously, theorists in these fields have relied on mathematical models (e.g. stochastic differential equations) based on fundamental assumptions borrowed from classical physics. What are these fundamental assumptions? Are they overly restrictive? Social and behavioral scientists also face findings that remain paradoxical from a classic probability point of view. These paradoxes suggest that measurements in these fields may not always obey the law of total probability and entail different assumptions. This program of research also will contribute to the training of students at the undergraduate, graduate, and post doctoral levels at two major state universities. In addition to student training, the investigators will make an effort to train scientists in the area quantum cognition. They have conducted a full day tutorial at the annual Cognitive Science meeting and they plan to continue these tutorials in the future. They also plan to organize a special issue on Quantum Cognition in the Journal of Mathematical Psychology. New graduate courses on quantum cognition and decision making will be prepared and presented at the graduate level, and finally, a resource web site will be developed with tutorial and reference information on quantum theory for social and behavioral sciences.

Project Summary

The objective of this research is to develop new neural network structures to solve optimal control problems with dynamic decision making. These problems are quite complex since the system dynamics could switch modes at unknown times based on event based decision making. The approach is to develop the decision-making paradigms from cognitive science principles but their mathematical representations will use Decision Field Theory. Their solutions contained in neural networks will interact with another set of networks that embed solutions to the related optimal control problem formulated in an approximate dynamic programming framework.

Intellectual Merit

This research seeks to find unified controller solutions to problems which have both continuous and discrete elements in them. It is expected that the mathematical cognitive science ideas developed will lead to new representations and problem solving structures in computational neuroscience and control. The work proposed in this effort seeks to accomplish these objectives by offering a transformative approach that integrates concepts from system science and cognitive science.

Broader Impact

Abstractions and solution structures developed through this research can be used in consequence or emergency management systems like managing the aftermath of an earthquake, retrieving an impaired aircraft to stability and sustainable motion and landing, and managing multiple assets and allocation in striking responses to threats. Decision making structures resulting from this research can make tremendous impact on human-machine interactions too. For example, driver aid systems can be developed to augment human perception and enhance their cognition when they drive under impaired conditions.

DESCRIPTION (provided by applicant): The purpose of this proposal is to extend our past work comparing performance of brain damaged, drug abusing, or psychopathological individuals and with non abusing or normal individuals on standard laboratory decision making tasks. Performance on these tasks is an interaction and synthesis of three different underlying components, including motivational, learning, and choice processes. Cognitive models of these complex decision tasks are used to break performance down into these three components. The parameters associated with these components are then used to understand the source of the decision making deficits exhibited by these clinical populations. Two critical assumptions underlying this past work are the assumptions of model generalization and parameter consistency. A model generalizes if one can fit the parameters of the model to one task for an individual, and then use these same parameters to predict performance on other closely related tasks for the same individual. Parameters are consistent if the parameters estimated from one task for an individual correlate with the parameters estimated from another closely related task for the same individual. These assumptions are crucial if we want to interpret the parameters as measuring stable characteristics of an individual, rather than some inessential characteristics of a laboratory task. So far, we achieved some initial success obtaining model generalization and parameter consistency. But success has been limited for at least two reasons: one is the need to find better models through model comparison, and the other is the need for better methods of estimating model parameters. We plan to improve our methods using new hierarchical Bayesian analyses. This new methodology allows one to build a model for individual differences rather than fitting individuals separately. This way the parameters for a single individual are estimated through a model which is informed by data from all individuals. This provides more stable parameter estimates and more powerful methods for model comparison. We also plan to extend the hierarchical Bayesian method for comparing model generalization.

This research project will develop and test a new measurement model based on quantum probability theory called the Hilbert space multi-dimensional model. With the striking advancement of modern data-collection methods, complex and massive data sets are generated from various sources and contexts that are conceptually connected. This promises to provide a better understanding of complex social and behavioral phenomena, but it also presents significant challenges for the integration and interpretation of data from multiple sources. The general Hilbert space multi-dimensional model will improve understanding of complex social and behavioral phenomena ranging from violations of rational decision theory to social survey data integration and interpretation. This project is part of a larger research program to build probabilistic and dynamic systems for social and behavioral sciences from quantum rather than classical probability principles. The project will develop and disseminate from public repositories self-contained software packages for applying and estimating the general Hilbert space multi-dimensional model in MATLAB, R, and Python.

The investigators will develop and test the general Hilbert space multi-dimensional model, including the development of the mathematical theory of the model and related statistical and computational tools for applying the model. They will rigorously test the model using a large range of experiments. When large data sets are collected from different contexts or conditions, often they can be summarized by contingency tables. A critical problem arises, however, regarding how to integrate and synthesize these tables into a compressed, coherent, and interpretable representation. A common solution is to try to construct a joint probability distribution to reproduce the frequency data observed in the tables. Bayesian causal networks then are often used to reduce the number of estimated parameters by imposing conditional independence assumptions. In many cases, however, no such joint distribution exists that can reproduce the observed tables. The general Hilbert space multi-dimensional model provides a promising solution to the problems faced by complex and massive data by constructing a single finite state vector that lies within a low dimensional Hilbert space and by forming a set of non-commuting measurement operators that represent the measurements. In this way, the model produces a compressed, coherent, and interpretable representation of the measured variables that form the complex collection of data tables even when no standard joint distribution exists.