An experiment was carried out to investigate the effects of magnitude of reward and consummatory time on the choice behaviour of human subjects. Analysis of the data revealed: (1) choice behaviour was a function of both magnitude of reward and consummatory time; (2) the effects of these two variables differed depending upon the period of training under consideration; (3) short consummatory times resulted in highly variable behaviour. The results were discussed in relation to theories derived from experiments using infrahuman subjects. / Thesis / Master of Arts (MA)

No description available.

Artificial intelligence (AI) has made astonishing breakthroughs in recent years and achieved comparable or even better performance compared to humans on many real-world tasks and applications. However, it is still far from reaching human-level intelligence in many ways. Specifically, although AI may take inspiration from neuroscience and cognitive psychology, it is dramatically different from humans in both what it learns and how it learns. Given that current AI cannot learn as effectively and efficiently as humans do, a natural solution is analyzing human learning processes and projecting them into AI design. This dissertation presents three studies that examined cognitive theories and established frameworks to integrate crucial human cognitive learning elements into AI algorithms to build human learning–augmented AI in the context of text analytics. The first study examined compositionality—how information is decomposed into small pieces, which are then recomposed to generate larger pieces of information. Compositionality is considered as a fundamental cognitive process, and also one of the best explanations for humans' quick learning abilities. Thus, integrating compositionality, which AI has not yet mastered, could potentially improve its learning performance. By focusing on text analytics, we first examined three levels of compositionality that can be captured in language. We then adopted design science paradigms to integrate these three types of compositionality into a deep learning model to build a unified learning framework. Lastly, we extensively evaluated the design on a series of text analytics tasks and confirmed its superiority in improving AI's learning effectiveness and efficiency. The second study focused on transfer learning, a core process in human learning. People can efficiently and effectively use knowledge learned previously to solve new problems. Although transfer learning has been extensively studied in AI research and is often a standard procedure in building machine learning models, existing techniques are not able to transfer knowledge as effectively and efficiently as humans. To solve this problem, we first drew on the theory of transfer learning to analyze the human transfer learning process and identify the key elements that elude AI. Then, following the design science paradigm, a novel transfer learning framework was proposed to explicitly capture these cognitive elements. Finally, we assessed the design artifact's capability to improve transfer learning performance and validated that our proposed framework outperforms state-of-the-art approaches on a broad set of text analytics tasks. The two studies above researched knowledge composition and knowledge transfer, while the third study directly addressed knowledge itself by focusing on knowledge structure, retrieval, and utilization processes. We identified that despite the great progress achieved by current knowledge-aware AI algorithms, they are not dealing with complex knowledge in a way that is consistent with how humans manage knowledge. Grounded in schema theory, we proposed a new design framework to enable AI-based text analytics algorithms to retrieve and utilize knowledge in a more human-like way. We confirmed that our framework outperformed current knowledge-based algorithms by large margins with strong robustness. In addition, we evaluated more intricately the efficacy of each of the key design elements. / Doctor of Philosophy / This dissertation presents three studies that examined cognitive theories and established frameworks to integrate crucial human cognitive learning elements into artificial intelligence (AI) algorithm designs to build human learning–augmented AI in the context of text analytics. The first study examined compositionality—how information is decomposed into small pieces, which are then recomposed to generate larger pieces of information. Design science research methodology has been adopted to propose a novel deep learning–based framework that can incorporate three levels of compositionality in language with significantly improved learning performance on a series of text analytics tasks. The second study went beyond that basic element and focused on transfer learning—how humans can efficiently and effectively use knowledge learned previously to solve new problems. Our novel transfer learning framework, which is grounded in the theory of transfer learning, has been validated on a broad set of text analytics tasks with improved learning effectiveness and efficiency. Finally, the third study directly addressed knowledge itself by focusing on knowledge structure, retrieval, and utilization processes. We drew on schema theory and proposed a new design framework to enable AI-based text analytics algorithms to retrieve and utilize knowledge in a more human-like way. Lastly, we confirmed our design's superiority in dealing with knowledge on several common text analytics tasks compared to existing knowledge-based algorithms.

Humans learn to interact with many complex dynamic systems such as helicopters, bicycles, and automobiles. This dissertation develops a subsystem identification method to model the control strategies that human subjects use in experiments where they interact with dynamic systems. This work provides new results on the control strategies that humans learn. We present a novel subsystem identification algorithm, which can identify unknown linear time-invariant feedback and feedforward subsystems interconnected with a known linear time-invariant subsystem. These subsystem identification algorithms are analyzed in the cases of noiseless and noisy data. We present results from human-in-the-loop experiments, where human subjects in- teract with a dynamic system multiple times over several days. Each subject’s control behavior is assumed to have feedforward (or anticipatory) and feedback (or reactive) components, and is modeled using experimental data and the new subsystem identifi- cation algorithms. The best-fit models of the subjects’ behavior suggest that humans learn to control dynamic systems by approximating the inverse of the dynamic system in feedforward. This observation supports the internal model hypothesis in neuro- science. We also examine the impact of system zeros on a human’s ability to control a dynamic system, and on the control strategies that humans employ.

Humans learn to interact with many complex physical systems. For example, humans learn to fly aircraft, operate drones, and drive automobiles. We present results from human-in-the-loop (HITL) experiments, where human subjects interact with dynamic systems while performing command-following tasks multiple times over a one-week period. We use a new subsystem identification (SSID) algorithm to estimate the control strategies (feedforward, feedforward delay, feedback, and feedback delay) that human subjects use during their trials. We use experimental and SSID results to examine the effects of system characteristics (e.g., system zeros, relative degree, system order, phase lag, time delay), reference command, and command-following objectives on humans command-following performance and on the control strategies that the humans learn. Results suggest that nonminimum-phase zeros, relative degree, phase lag, and time delay tend to make dynamic systems difficult for human to control. Subjects can generalize their control strategies from one task to another and use prediction of the reference command to improve their command-following performance. However, this dissertation also provides evidence that humans can learn to improve performance without prediction. This dissertation also presents a new SSID algorithm to model the control strategies that human subjects use in HITL experiments where they interact with dynamic systems. This SSID algorithm uses a two-candidate-pool multi-convex-optimization approach to identify feedback-and-feedforward subsystems with time delay that are interconnected in closed loop with a known subsystem. This SSID method is used to analyze the human control behavior in the HITL experiments discussed above.

This dissertation explores workforce planning in manufacturing and healthcare systems. In manufacturing systems, the existing workforce planning models often lack fidelity with respect to the mechanism of learning. Learning refers to that employees’ productivity increases as they gain more experience. Workforce scheduling in the short term has a longer term impact on organizations’ capacity. The mathematical representations of learning are usually nonlinear. This nonlinearity complicates the planning models and provides opportunities to develop solution methodologies for realistically-sized instances. This research formulates the workforce planning problem as a mixed-integer nonlinear program (MINLP) and overcomes the limitations of cur- rent solution methods. Specifically, this research develops a reformulation technique that converts the MINLP to a mixed integer linear program (MILP) and proposes several techniques to speed up the solution time of solving the MILP. In organizations that use group work, workers learn not only by individual learning but also from knowledge transferred from team members. Managers face the decision of how to pair or team workers such that organizations benefit from this transfer of learning. Using a mathematical representation that incorporates both in- dividual learning and knowledge transfer between workers, this research considers the problem of grouping workers to teams and assigning teams to sets of jobs based on workers’ learning and knowledge transfer characteristics. This study builds a Mixed- integer nonlinear programs (MINP) for parallel systems with the objective of maximizing the system throughput and propose exact and heuristic solution approaches for solving the MINLP. In healthcare systems, we focus on managing medical technicians in medical laboratories, in particular, the phlebotomists. Phlebotomists draw specimens from patients based on doctors’ orders, which arrive randomly in a day. According to the literature, optimizing scheduling and routing in hospital laboratories has not been regarded as a necessity for laboratory management. This study is motivated by a real case at University of Iowa Hospital and Clinics, where there is a team of phlebotomists that cannot fulfill doctors requests in the morning shift. The goal of this research is routing these phlebotomists to patient units such that as many orders as possible are fulfilled during the shift. The problem is a team orienteering problem with stochastic rewards and service times. This research develops an a priori approach which applies a variable neighborhood search heuristic algorithm that improves the daily performance compared to the hospital practice.

Modeling comprehensive human decision behaviors in a unified and extensible framework is quite challenging. In this research, an integrated Belief-Desire-Intention (BDI) modeling framework is proposed to represent the human decision behavior, whose submodules (Belief, Desire, Decision-Making, and Emotion modules) are based on a Bayesian belief network (BBN), Decision-Field-Theory (DFT), a probabilistic depth first search (PDFS) technique, and a BBN-reinforcement (Q-Learning) hybrid learning algorithm. A key novelty of the proposed model is its ability to represent various human decision behaviors such as decision-making, decision-planning, and learning in a unified framework.To this end, first, we extend DFT (a widely known psychological model for preference evolution) to cope with dynamic environments. The extended DFT (EDFT) updates the subjective evaluation for the alternatives and the attention weights on the attributes via BBN under the dynamic environment. To illustrate and validate the proposed EDFT, a human-in-the-loop experiment is conducted for a virtual stock market. Second, a new approach to represent learning (a dynamic evolution process of underlying modules) in the human decision behavior is proposed under the context of the BDI framework. Our research focuses on how a human adjusts his perception process (involving BBN) dynamically against his performance (depicted via a confidence index) in predicting the environment as part of his decision-planning. To this end, Q-learning is employed and further developed.To mimic realistic human behaviors, attributes of the BDI framework are reverse-engineered from human-in-the-loop experiments conducted in the Cave Automatic Virtual Environment (CAVE). The proposed modeling framework is demonstrated for a human's evacuation behaviors in response to a terrorist bomb attack. The constructed simulation has been used to test the impact of several factors (e.g., demographics, number of police officers, information sharing via speakers) on evacuation performance (e.g., average evacuation time, percentage of casualties).In addition, the proposed human decision behavior model is extended for decisions of many stakeholders that form a complex social network in the community-based development of software systems.To the best of our knowledge, the proposed human decision behavior modeling framework is one of the first efforts to represent various human decision behaviors (e.g., decision-making, decision-planning, dynamic learning) in a unified BDI framework.

Relational Density Theory describes quantifiable higher-order properties governing relational framing of verbal organisms. Consistent with Newtonian classical mechanics, the theory posits that relational networks, and relating itself, will demonstrate the higher-order emergent properties of density, volume, and mass. Thus, networks that contain more relations (volume) that are stronger (density) will be more resistant to change (i.e., contain greater mass; mass = volume * density). Consistent with Newton’s law of gravity, networks that contain greater mass will also demonstrate force, accelerating the acquisition of new relations beyond that accounted for by direct acting contingencies, therefore demonstrating emergent self-organization that is highly susceptible to small changes at initial conditions. The current set of experiments provides initial proof of concept data for foundational principles introduced in the theory. Experiment 1 (N = 6) models the volumetric mass density formula, predicting that networks with greater volume and density will be more resistant to change (i.e., contain greater mass) when counterconditioning is applied to a subset of derived relations contained within experimentally established networks. Results were consistent with theoretical predictions based on density on 10 of 12 occasions, and resistance appeared greater for relations operating at greater volume. Experiment 2 (N = 6) extended directly from Experiment 1, generating a density differential through exposure at initial training conditions, and utilizing response time as a measure of relational density. Results again demonstrated successful prediction of resistance corresponding with the emergent density differential on 10 of 12 occasions, along with overall greater resistance corresponding with and volumetric increases. Experiment 3 (N = 9) demonstrated that relational volume can detract from relational density when accurate responding is near 100%, and that network density is predictive of class mergers when no merged responding is ever reinforced, suggesting that network mass can exert force on relational responding in the absence of any experimental conditioning (i.e., gravity). Taken together, results have radical implications for understanding the self-emergent nature of complex human behavior, with applications in therapy and treatment, as well as in understanding the human condition more broadly.

Traffic conflicts associated to signalized intersections are one of the major contributing factors to crash occurrences. Driver behavior plays an important role in the safety concerns related to signalized intersections. In this research effort, dynamics of driver behavior in relation to the traffic conflicts occurring at the onset of yellow is investigated. The area ahead of intersections in which drivers encounter a dilemma to pass through or stop when the yellow light commences is called Dilemma Zone (DZ). Several DZ-protection algorithms and advance signal settings have been developed to accommodate the DZ-related safety concerns. The focus of this study is on drivers' decision dynamics, human learning, and choice behavior in DZ, and DZ-related safety measures. First, influential factors to drivers' decision in DZ were determined using a driver behavior survey. This information was applied to design an adaptive experiment in a driving simulator study. Scenarios in the experimental design are aimed at capturing drivers learning process while experiencing safe and unsafe signal settings. The result of the experiment revealed that drivers do learn from some of their experience. However, this learning process led into a higher level of risk aversion behavior. Therefore, DZ-protection algorithms, independent of their approach, should not have any concerns regarding drivers learning effect on their protection procedure. Next, the possibility of predicting drivers' decision in different time frames using different datasets was examined. The results showed a promising prediction model if the data collection period is assumed 3 seconds after yellow. The prediction model serves advance signal protection algorithms to make more intelligent decisions. In the next step, a novel Surrogate Safety Number (SSN) was introduced based on the concept of time to collision. This measure is applicable to evaluate different DZ-protection algorithms regardless of their embedded methodology, and it has the potential to be used in developing new DZ-protection algorithms. Last, an agent-based human learning model was developed integrating machine learning and human learning techniques. An abstracted model of human memory and cognitive structure was used to model agent's behavior and learning. The model was applied to DZ decision making process, and agents were trained using the driver simulator data. The human learning model resulted in lower and faster-merging errors in mimicking drivers' behavior comparing to a pure machine learning technique. / Ph. D.

Brain-computer interfaces (BCIs) create a new form of communication and control for humans by translating brain activity directly into actions performed by a computer. This new field of research, best known for its breakthroughs in enabling fully paralyzed or locked-in patients to communicate and control simple devices, has resulted in a variety of remarkable technological developments. However, the field is still in its infancy, and facilitating control of a computer application via thought in a broader context involves a number of a challenges that have not yet been met. Advancing BCIs beyond the experimental phase continues to be a struggle. End-users have rarely been reached, except for in the case of a few highly specialized applications which require continual involvement of BCI experts. While these applications are profoundly beneficial for the patients they serve, the potential for BCIs is much broader in scope and powerful in effect. Unfortunately, the current approaches to brain-computer interfacing research have not been able to address the primary limitations in the field: the poor reliability of most BCIs and the highly variable performance across individuals. In addition to this, the modes of control available to users tend to be restrictive and unintuitive (\emph{e.g.}, imagining complex motor activities to answer ``Yes" or ``No" questions). This thesis presents a novel approach that addresses both of these limitations simultaneously. Brain-computer interfacing is currently viewed primarily as a machine learning problem, wherein the computer must learn the patterns of brain activity associated with a user's mental commands. In order to simplify this problem, researchers often restrict mental commands to those which are well characterized and easily distinguishable based on \emph{a priori} knowledge about their corresponding neural correlates. However, this approach does not fully recognize two properties of a BCI which makes it unique to other human-computer interfaces. First, individuals can vary widely with respect to the patterns of activation associated with how their brains generate similar mental activity and with respect to which kinds of mental activity have been most trained due to life experience. Thus, it is not surprising that BCIs based on predefined neural correlates perform inconsistently for different users. Second, for a BCI to perform well, the human and the computer must become a cohesive unit such that the computer can adapt as the user's brain naturally changes over time and while the user learns to make their mental commands more consistent and distinguishable given feedback from the computer. This not only implies that BCI use is a skill that must be developed, honed, and maintained in relation to the computer's algorithms, but that the human is the fundamental component of the system in a way that makes human learning just as important as machine learning. In this thesis it is proposed that, in the long term, a generalized BCI that can discover the appropriate neural correlates of individualized mental commands is preferable to the traditional approach. Generalization across mental strategies allows each individual to make better use of their own experience and cognitive abilities in order to interact with BCIs in a more free and intuitive way. It is further argued that in addition to generalization, it is necessary to develop improved training protocols respecting the potential of the user to learn to effectively modulate their own brain activity for BCI use. It is shown through a series of studies exploring generalized BCI methods, the influence of prior non-BCI training on BCI performance, and novel methods for training individuals to control their own brain activity, that this new approach based on balancing the roles of the user and the computer according to their respective capabilities is a promising avenue for advancing brain-computer interfacing towards a broader array of applications usable by the general population. / Thesis / Doctor of Philosophy (PhD)