A Closer Look at
Exploratory Learning of Interactive Devices
This paper describes a new rational framework for modelling exploratory learning of interactive devices, first presented in Young & Cox (2000). The framework has been used as a basis for analysing a number of protocols taken from participants exploring a simulated central heating timer that provide examples supporting the framework. The results suggest that the framework can successfully explain episodes of the participants' behaviour during exploration and also what they learn from it. Finally, we describe the modeling work (in progress) in Soar (Newell, 1990) and COGENT (Cooper & Fox, 1998) implementing the framework.
Socionics: A New
Challenge for Cognitive Modeling
At the line of intersection between Distributed Artificial Intelligence, Multi-Agent-Based Modeling and Sociology a new research field has emerged: Socionics. Main aim of this new research discipline is to build and study artificial societies and to develop intelligent computer technologies by picking up sociological concepts. This paper attempts to show why Cognitive Modeling (CM) should also pick up this challenge.
Strategies in a
Complex Game and their Background
We tried to simulate the behaviour of 30 Ss in a complex
game situations. It turned out that the variance of human Ss was much higher
with respect to many behavioural measures than the variance of simulated Ss.
This is due to the fact that human Ss use more different strategies than the
computer did and change strategies deliberately. On the one hand the choice of
a strategy seems to be dependent on different "emotional" parameters
of the Ss. One important parameter is for instance the stability of self-efficacy.
Simulating emotional parameters yields an increase of the strategic variety of
the simulated Ss and hence an increase of the similarity of human and simulated
behaviour. On the other hand phases of self-reflection, which can be found with
some Ss, can be the reason for the choice and especially for the change of a
strategy.
Facial Expression
Recognition with Modular Neural Networks
We construct a modular neural network to perform facial
expression recognition. The structure is a 4 layer feed-forward architecture
consisting of two parts: first, a hidden layer with unsupervised learning is applied
to the input images to obtain a reduced representation of the image; second,
the network splits into modules ("experts") specialized in the
different expressions, modules that are trained through the back-propagation
algorithm. Using the Yale Face database for 14 subjects displaying 4 emotions
each (neutral, happy, sad and surprise faces) we obtain a generalization rate
of 78.6 % on unseen faces using a hebbian unsupervised learning scheme compared
to 70.8 % when using random weights
Human-Task
Adaptations: The Next Step for Cognitive Modeling
Modeling Cognitive
versus Perceptual-Motor Tradeoffs using ACT-R/PM
Information stored in-the-world is retrieved from external memory via visual perception as rendered by the appropriate saccades and fixations. Recent research has suggested that when information in-the-world is readily accessible, internal storage is not needed. Perceptual-motor strategies will be deployed to reacquire information as needed. However, Fu & Gray (2000) found that when the cost of information access was increased from a simple key press to a one-second lockout time, the perceptual-motor strategy was replaced with a strategy that placed task-relevant information into working memory. This suggests that the decision to store information in-the-head versus in-the-world is sensitive to effort considerations. In this paper, we present our work-in-progress report on modeling the data using ACT-R/PM.
Failure to Learn
from Negative Feedback in a Hierarchical Adaptive System
A simulation of hierarchical learning shows that negative feedback is ineffective when it influences the choice of high-level goals before the choice of subordinate low-level actions. Attention to errors committed while perfecting the execution of an appropriate high-level goal may lead to the incorrect inference that the goal is inappropriate.
The Strategic Use
of Memory for Frequency and Recency in Search Control
A requirement of an information processing account of human problem solving is that it includes a mechanism by which people remember which goals and operators have been evaluated and which still need to be evaluated. This is an issue that has been glossed in the two major architectural accounts of cognition. We report a model of human search that depends on independent estimates of frequency and recency in order to control search.
An ACT-R Model of
Syllogistic Inference
We describe SYLLOG, a model of syllogistic inference, built within the ACT-R framework. Its construction was guided by data obtained from 88 subjects performing the syllogistic inference task. The model’s inference engine uses the PC inference rule Hypothetical Syllogism (HS) and a set of modal logic transformation rules either to create a representation of the premises that allows an appropriate use of HS, or to check the appropriateness of having already used HS to obtain a putative conclusion.
Modeling an
Opportunistic Strategy for Information Navigation
A computational model of a user navigating Web pages was used to identify factors that affect Web site usability. The model approximates a typical user searching for specified target information in architectures of varying menu depth. Search strategies, link ambiguity, and memory capacity were varied and model predictions compared to human user data. A good fit to observed data was obtained for a model that assumed users 1) used little memory capacity; 2) selected a link whenever its perceived likelihood of success exceeded a threshold; and, 3) opportunistically searched below threshold links on selected pages prior to returning to the parent page.
Interactions
between Frequency and Age of Acquisition
The performance of a connectionist network, in which some resources are absent or damaged, is examined as a function of various learning parameters. The robustness of each learned item is shown to be a function of the time at which it was "acquired" by the network and its overall frequency in the environment. Acquisition time of each item and its frequency are both shown to influence robustness under several learning conditions.
On the Normativity
of Failing to Recall Valid Advice
Instructed category learning tasks involve the acquisition of a categorization skill from two sources of information: explicit rules provided by a knowledgeable teacher and experience with a collection of labeled examples. Studies of human performance on such tasks have shown that practice categorizing a collection of training examples can actually interfere with the proper application of explicitly provided rules to novel items. In this work, the normativity of such exemplar-based interference is assessed by confronting a model of optimal memory performance with such a task and comparing the "rational" model's behavior with that exhibited by human learners. When augmented with a rehearsal mechanism, this optimal memory model is shown to match human responding, producing exemplar-based interference by relying on memories of similar training set exemplars to categorize a novel item, in favor of recalling rule instructions.
The Grain Size of
Cognitive Models: How Low Should We Go?
There has been vigorous debate regarding the level at which it is appropriate and/or necessary to model human cognition (Smolensky, 1988). It is difficult to settle this issue because the many unknowns regarding human cognition confound efforts to attribute the success or failure of any particular modeling effort to one factor alone. In order to study this matter without a large number of confounds, it is useful to consider how grain size affects the predictive power of computational models in physical domains that are simpler than cognition. For this purpose, a model of a physical system that is simple to describe but nonetheless has chaotic behavior has been used to isolate grain size as a variable that affects the behavior of the model in clear ways. While this does not settle the question of how grain size affects the accuracy of cognitive models, it does offer several possibilities for how grain size may in principle affect the accuracy of a model. This has one advantage over debate that is concerned solely with the domain of primary interest (cognition) but cannot arrange any clean tests of its hypotheses.
Human Performance
in Transverse patterning with a Hippocampal Model
The hippocampus is necessary in humans and rats for tasks that require both elemental and conjunctive associations, such as transverse patterning. This study applies a biologically plausible model of the CA3 region of the hippocampus to human data on this task. The analysis demonstrates how neural codes can represent input items in their spatio-temporal context, thereby forming both elemental and conjunctive representations.
Modeling Behavior
in Complex and Dynamic Situations - The Example of Flying an Automated Aircraft
In basic research, cognitive modeling has proven a valuable methodology for explicating theoretical assumptions, testing their dynamic interactions, and exploring the scope of theories. Cognitive architectures such as ACT-R or Soar provide a common basis for different models and enhance communication and exchange of solutions. In applied contexts too, modeling of real tasks and operators could further the understanding of human-machine systems; validated models could provide an objective guide to design and training decisions. However, as real tasks typically require more knowledge and are more complex than laboratory tasks, content independent cognitive architectures do not sufficiently constrain the modeling of these tasks. We argue that - despite this problem - models of behavior in real world tasks should also be developed within established architectures. In doing so, it is important to specify what parts of the model are derived directly from the architecture and for wha! t parts new solutions had to be developed. Thus, models benefit from the broad empirical confirmation of the architecture, which in turn benefits from the identification of domains where it needs to be extended. As an example for this approach, we present an ACT-R model that simulates the interaction between airline pilots and the flight management system. Most aspects of the task for which ACT-R didn't provide enough constraints followed from the task's dynamic and the requirement to interleave subtasks during long time intervals. Our research yielded valuable hints how the scope of ACT-R could be extended to reasoning and action in more complex and dynamic environments.
ACT-RS: A
Neuropsychologically Inspired Module for Spatial Reasoning
We present an extension to ACT-R, called ACT-RS, which adds a neurologically-inspired module for representing and interacting with space. The module includes four functionally different representations of external space that vary along dimensions of input modalities, tasks supported, format, resolution, extent, and features encoded.
Spatial Navigation
Using Hierarchical Cognitive Maps
Planning point-to-point paths between distant locations in a large environment can become a grueling task. The problem results from the limited amount of information that an agent can process at a given time in working memory. This work describes an agent capable of building and using a hierarchical cognitive map in a large-scale environment. Computer simulations show that the agent is able to guide exploration, and to plan paths between places.
Can Cognitive
Modeling Improve Usability Testing and Rapid Prototyping?
We argue that usability testing, as employed in the rapid prototyping cycle, could be improved by testing simulated users. We briefly review some of the benefits that this methodology could offer and discuss one approach to building a simulated user using ACT-R to embody a GOMS-like memory structure.
Modeling the
Inverted-U Effect with ACT-R
‹
What People Learn
from Exploratory Device Learning
The empirical investigations outlined in this paper are concerned with identifying what people do during exploratory learning and what they learn whilst they conduct it. The results have been used to develop a framework within which to model exploration. Implementation of the framework is proposed in order to test a number of predictions and to help explain and support findings from further empirical studies.
The influence of
experienced effort on learning and choice of solution paths in a simple
navigation task
In my dissertation, I am going to study the influence of experienced effort on learning and problem-solving behavior. In a simple navigation task on a computer-simulated map, subjects have to acquire information on various levels of effort in different solution paths through experience. The experienced effort information allows subjects to improve their performance by finding faster solution paths on the map. I am planning to build ACT-R model to understand the underlying mechanisms. Specifically, I am interested in modeling the learning of the "current" effort and the "downstream" effort in ACT-R theory, and how each of them can influence problem-solving behavior in the task.
Plural Morphology
in Compounding is not Good Evidence to Support the Dual Mechanism model
The compounding phenomenon is considered to be good evidence to support the dual mechanism model of morphological processing (Pinker & Prince, 1992). However evidence from initial neural net modeling has shown that a single route associative memory based account might provide an equally, if not more valid explanation of the treatment of plurals in compounds. Further neural net modeling and empirical work is proposed to test this single route account.
Modeling User
Knowledge and Semantic Structure for Information Extraction from Text
Latent Semantic Analysis (LSA, Landauer & Dumais, 1997) is used to represent user knowledge and to extract user relevant semantic structures from text. A model of user knowledge is created and empirically optimized. This model is then used to extract user relevant semantic structure from text. User model and quality of extracted semantic structure are empirically evaluated.
Statistical
Learning of Human Faces
SHRUTI-agent: a
structured connectionist model of decision-making
A neurally plausible connectionist model of decision-making, based on the SHRUTI architecture, is being de-vloped. Toward this end, issues of appropriate connec-tionist representations for belief and utility, necessary control mechanisms, and reinforcement-based learning are addressed.