1,720,992 research outputs found
Cognitive Model for a Mental Rotation Task
No full textBespoke cognitive models of mental spatial transformation, like those used in mental rotation tasks, can generate a very close fit to human data. However these models usually lack grounding to a common spatial theory. In turn, this makes it difficult to assess their validity and impedes research insights that go beyond task specific limitations. We introduce a spatial module for the cognitive architecture ACT-R, serving as a framework offering unified mechanisms for mental spatial transformation to try and alleviate those problems. This module combines symbolic and spatial information processing for three-dimensional objects, while suggesting constraints on this processing to ensure high theoretical validity and cognitive plausibility. A mental rotation model was created to make use of this module, avoiding custom-made mechanisms in favor of a generalizable approach. Results of a mental rotation experiment are reproduced well by the model, including effects of rotation disparity and improvement over time on reaction times. Based on this, the spatial module might serve as a stepping stone towards unified, application oriented research into mental spatial transformation.DFG, 396560184, Digitales Produkt - Digitaler Nutzer (DPDN
Cognitive Model for a Mental Folding Task (2023)
No full textA simulated replication of the experiment was created for a cognitive model solving the mental folding task. The model was formalised and implemented in ACT-R, based on cognitive processes theorised by Shepard and Feng [1972]. Like the mental rotation model, it followed an outline proposed by Just and Carpenter [1976] comprised of visual encoding, transformation and comparison, and motor response. The mental folding model held several differences to its mental rotation counterpart. Contrary to the rotation model, no two separate strategy tracks were assumed by the folding model, making it more deterministic. Secondly, instead of alternating between comparison and transformation during folding processes, the mental folding model completed both necessary transformations (one for each arrow square), and afterwards initiated a single comparison process of the resulting structure to the reference stimulus. Akin to the mental rotation model however, instance-based learning [Gonzalez et al., 2003] was used to allow the model to skip the transformation process, thereby potentially shortening trial time: if the model recognised a target stimulus as being solved before and could remember its associated solution, spatial transformation was bypassed. In addition, visual shortcuts were possible if certain patterns were recognised (e.g., one arrow pointing to an empty square on one stimulus but not the other, or geometric relations between squares) and could be decided on by the model, with probability of choice mediated by a reinforcement learning algorithm included in ACT-R [Fu and Anderson, 2004].
The updated model focusses on modeling the data reported by Hilton et al., 2021, and contains bugfixes.DFG, 396560184, Digitales Produkt - Digitaler Nutzer (DPDN
Cognitive Model for a Mental Rotation Task (2023)
No full textThe cognitive model solved a simulated version of the mental rotation task provided to the participants, based on cognitive processes theorised by Shepard and Metzler [1971], Cooper and Shepard [1973], Just and Carpenter [1976], and Yuille and Steiger [1982]. It followed the basic processing stages (visual encoding, transformation and comparison, and motor response) described by Cooper and Shepard [1973] and Just and Carpenter [1976]. Two separate solving tracks were assumed by the model, akin to the wholesale and piecemeal solving strategies originally suggested by Yuille and Steiger [1982]. An instant-based learning mechanism included in ACT-R [Gonzalez et al., 2003] is used to choose the appropriate strategy based on familiarity with the presented stimulus.
The updated model is set to simulate activity data to compare to dipole-modeled and clustered EEG data from Hilton et al. (2021), and contains small bugfixes.DFG, 396560184, Digitales Produkt - Digitaler Nutzer (DPDN
Spatial Module for ACT-R (1.4)
No full textThe spatial module extends ACT-R's modular structure by a dedicated processing unit for mental spatial transformations. Based on work by Gunzelmann & Lyon (2007), the idea is to offer functionality for processing three-dimensional data in ACT-R in a cognitively plausible fashion. In contrast to their work, the module presented herein avoids episodic and allocentric buffers, as I perceive the mechanisms underlying these additional buffers to be already supplied by ACT-R's default modules, namely the declarative and visual/imaginal modules.
The spatial module aims to offer better explainability, applicability and validity for cognitive models of spatial cognition by offering explanations for commonly shown effects such as differences in spatial strategies or increased solving time for higher task difficulties, supporting multiple paradigms of mental spatial cognition research such as mental rotation or mental folding and offering a common framework for these mechanisms, thus bypassing the need for overly specific modeling approaches.
This version of the spatial module fixes an issue with reset commands that could occur when model runs were cancelled, ran out of time or otherwise left the spatial or spatial-action buffers unresolved.DFG, 396560184, Digitales Produkt - Digitaler Nutzer (DPDN
Nutzendenerfahrung erfassen und verstehen zur Verbesserung der Mensch-Roboter-Interaktion
No full textAs robotic systems become increasingly integrated into everyday life, improving the user experience in human-robot interaction can facilitate their acceptance and long-term use.
This thesis examines how trust and embodiment, two key aspects of the user experience, influence non-expert users in two domains: adaptive collaboration with a robotic arm and embodiment in the robotic limb illusion. These factors are critical for user acceptance, shaping how users interact with robotic systems.
The first part investigates trust in adaptive collaborative robots. One approach to enhancing trust is human-centered reinforcement learning, particularly learning from demonstration, where users teach robots new tasks through direct interaction. Two user studies explore how adding human training data to reinforcement learning agents affects trust. The first study examines whether human-trained agents influence user trust, while the second extends this to a physical robotic arm, where participants train the robot themselves. Results show that incorporating expert human data enhances trust, even when performance improvements are not statistically significant.
The second part focuses on embodiment in the robotic limb illusion, building on research from the rubber hand illusion. Bayesian causal inference models are applied to analyze embodiment, first by fitting a model to robotic leg illusion data, showing that visual priors play a crucial role in multisensory integration. A follow-up user study gathers sufficient data for individual-level analysis, allowing the model to be adapted for personalized embodiment responses. By estimating user-specific priors, this study provides insights into individual embodiment differences and supports the design of personalized prosthetic and robotic augmentation devices.
Together, these studies contribute to human-centered human-robot interaction research by demonstrating how trust and embodiment can be systematically influenced through human data integration and cognitive modeling. The findings suggest that (1) trust in collaborative robots should be considered separately from task performance and (2) personalized embodiment models could enhance prosthetic and robotic limb integration. By addressing these challenges, this thesis advances the development of adaptive, user-friendly robotic systems, ensuring they align more closely with human needs and expectations
Data and Statistics for Preuss et al., 2024
No full textThe existence of a general mental spatial transformation mechanism in human cognition has long been argued over. Consensus exists on the importance of parietal areas for spatial transformation, however, lateralization, differentiation, and involvement of additional neural areas seem strongly dependent on task design, including conditions, context, and chronology. This study finds common ground and distinct features of spatial transformation in two tasks, mental rotation and mental folding, by analyzing and comparing EEG recordings from two experiments. Cognitive modeling was used to find brain areas associated with mental rotation and mental folding by linking simulation and cortical data through a novel approach. Task-specific models simulated intra-trial cognitive activity, predicting neural activity sources and providing theory-driven semantic interpretations of neural activity during task-solving. Mental rotation showed spatial activity in parietal and occipital areas, with central and right regions showing increased activity for easier trials and left regions for more difficult trials. For mental folding, the results showed central parietal and left parietal as well as occipital areas during spatial storage activity, as well as right parietal areas exclusive to spatial transformation. Left occipital and parietal regions were particularly active for visual baseline trials, while the right parietal area exhibited stronger activity for higher task difficulty. Comparing neural correlates between tasks showed inverse, difficulty-dependent lateralization patterns, implying contrasting demands on representation and transformation processes between rotation and folding.DFG, 396560184, Digital Product - Digital User (DPDU
Cognitive Model for a Combined Mental Rotation and Folding Task
No full textThe cognitive model is supplied with a simulated experiment replicating the original combined mental rotation and folding task design as presented to the participants. Due to the novelty of the task, its cognitive model is not directly based on existing literature, but uses aspects suggested for mental rotation and mental folding (Shepard & Metzler 1971, Shepard & Feng 1972, Just & Carpenter 1976, Yuille & Steiger 1982, Wright 2008) while relying on learning mechanisms implemented in ACT-R (Gonzales et al. 2003, Fu & Anderson 2004).
On reference stimulus onset, the base square with dot marking is located on the figure and its location relative to the square (top left, bottom left, bottom right or top right) encoded symbolically. In addition, the model determines if the presented figure is a two-dimensional folding pattern (no folding condition), or a three-dimensional partially folded cube. After target stimulus onset, the model will proceed with either a direct visual comparison (only for no folding trials), mental rotation, or mental folding (only for folding trials).
1) If visual comparison is chosen and establishes both figures as equal (exhibiting both no rotation and no folding, i.e. constituting baseline trials), the dot position on both stimuli is compared directly and a response is generated, pre-empting the need for spatial transformations. Otherwise, mental rotation is initiated.
2) If rotation is chosen, the target stimulus is encoded either piecemeal or wholesale, contingent on stimulus familiarity as decided by instance retrieval of the target figure outline. If retrieval is successful, the model is allowed to visuospatially encode all arms of the folding pattern at once, otherwise, each arm is encoded individually and appended to a combined structure (while piecemeal mental rotation usually refers to transforming individual pieces, we opted for an approach of subsequently merging arms into a single figure before transformation, as these often consist of a single square). Then, the spatially encoded structure is rotated sequentially and compared to the reference stimulus after each step, unless its rotation matches the reference. In case the reference figure is three-dimensional, the spatial target structure is statically rotated into the same perspective. If required, the model will continue with mental folding.
3) If folding is chosen, the arm of the target containing the dot marker is visually encoded as a spatial structure. An instance retrieval mechanism is started to look for known completed structures associated with the target pattern and, if successful, encodes the completed folded arm directly, thereby bypassing the transformation. If no instance is found, the arm will be sequentially folded by 90 degrees at each of its folding edges starting from the base square and moving towards the square containing the dot marker, until the latter is folded in its final position. If required, the model will continue with mental rotation.
After all necessary spatial transformations are completed, the dot marker position on the mental spatial structure is compared to the reference dot marker position, which is either visually encoded again or remembered from its initial appearance. Finally, a match or mismatch response is initiated per simulated motor response simulating a button press.DFG, 396560184, Digitales Produkt - Digitaler Nutzer (DPDN
Ermöglichen von Pilotenassistenz mithilfe eines kognitiven Zwillings
No full textPiloting large commercial airliners is a complex task that demands high levels of attention and precise decision-making, especially in non-routine situations. Delivering timely and appropriate support to pilots is challenging. To effectively assist pilots, an assistance system must embody four key capabilities: (1) contextual awareness, (2) sensitivity to individual pilot needs, (3) the ability to anticipate future pilot actions, and (4) the capacity to learn from experience and apply this knowledge to new situations.
This dissertation advances the development of pilot assistance systems by implementing and demonstrating these four capabilities through the construction of a cognitive twin. It comprises four scientific publications, each dedicated to one of these essential abilities. The work presented here introduces a model-based system that tracks and predicts pilot behavior during flight operations. The system builds upon several key methods: cognitive modeling using the ACT-R cognitive architecture, a technique from intelligent tutoring known as model tracing, integration with aircraft systems via recordings of aircraft, environment, and pilot action parameters, and the incorporation of (neuro-)physiological data from the pilot.
Two extensive empirical studies were conducted to develop and evaluate the cognitive twins presented in each publication. Additionally, the papers introduce and elaborate on innovative cognitive modeling techniques, including neuroadaptive cognitive modeling through model tracing and the integration of neurophysiological data, the combination of supervised learning with cognitive modeling, and the use of open retrievals in combination with instance-based learning.
The findings demonstrate that models which are context-aware, pilot-adapted, anticipatory, and capable of transfer learning significantly outperform traditional models in accurately representing pilots' responses to cockpit alerts and events. This research paves the way for reliable and responsive cognitive assistance systems in aviation.Das Führen großer Verkehrsflugzeuge ist eine komplexe Aufgabe, die ein hohes Maß an Aufmerksamkeit und präzise Entscheidungsfindung erfordert, insbesondere in nicht routinemäßigen Situationen. Dabei ist es nicht einfach, Piloten rechtzeitige und angemessene Unterstützung anzubieten. Um Piloten effektiv unterstützen zu können, muss ein Assistenzsystem vier Schlüsselfähigkeiten aufweisen: (1) Bewusstsein über Anforderungen des Kontexts, (2) Sensibilität für die individuellen Bedürfnisse von Piloten, (3) die Fähigkeit, künftige Pilotenhandlungen zu antizipieren, und (4) die Fähigkeit, aus Erfahrungen zu lernen und dieses Wissen auf neue Situationen anzuwenden.
Diese Dissertation treibt die Entwicklung von Pilotenassistenzsystemen voran, indem sie diese vier Fähigkeiten in der Form eines kognitiven Zwillings implementiert und demonstriert. Sie umfasst vier wissenschaftliche Veröffentlichungen, die sich jeweils mit einer der oben genannten essentiellen Fähigkeiten befassen. Die vorgestellte Arbeit stellt ein modellbasiertes System vor, das das Verhalten des Piloten während des Flugbetriebs verfolgt und vorhersagt. Dabei kommen übergreifend die folgenden Methoden zum Einsatz: kognitive Modellierung mit Hilfe der kognitiven Architektur ACT-R, eine Technik aus dem Bereich intelligenter tutorieller Systeme namens “Model Tracing”, die Integration mit Flugzeugsystemen über Aufzeichnungen Flugzeug-, Umgebungs- und Pilotenaktionsparametern, sowie die Einbeziehung von (neuro-)physiologischen Daten des Piloten.
Zur Entwicklung und Bewertung der in jeder Veröffentlichung vorgestellten kognitiven Zwillinge wurden zwei umfangreiche empirische Studien durchgeführt. Darüber hinaus werden in den Artikeln innovative Techniken zur kognitiven Modellierung entwickelt, angewandt und erläutert, darunter die neuroadaptive kognitive Modellierung mit Hilfe von “Model Tracing” und der Integration neurophysiologischer Daten, die Kombination von überwachtem Lernen mit kognitiver Modellierung und die Verwendung sogenannter “open retrievals” in Kombination mit instanzbasiertem Lernen.
Die Ergebnisse zeigen, dass Modelle, die kontextbewusst, an den Piloten angepasst, antizipierend und zum Transferlernen fähig sind, herkömmliche Modelle bei der genauen Abbildung von Pilotenreaktionen auf Warnungen und Ereignisse im Cockpit deutlich übertreffen. Diese Forschung ebnet den Weg für zuverlässige und reaktionsfähige kognitive Assistenzsysteme in der Luftfahrt
Towards a General Model of Repeated App Usage
No full textThe main challenge of implementing cognitive models for usability testing lies in reducing the modeling effort, while including all relevant cognitive mechanisms, such as learning and relearning, in the model. In this paper we introduce a general cognitive modeling approach with ACT-R for hierarchical, list-based smartphone apps. These apps support the task of selecting a target, via navigating through subtargets positioned on different layers. Mean target selection time for repeated app interaction, learning and relearning behavior was collected in four studies conducted with either a shopping app or a real-estate app. The predictions of the general modeling approach match the empirical data very well, both in terms of trends and absolute values. We also explain how such a general modeling approach can be followed. The presented general model approach requires little modeling effort to be used for predicting overall efficiency of other apps. It supports more complex interface, as well
Recommended from our members
Proposing a Cognitive System for Universal Mental Spatial Transformations
Get PDFMental spatial transformation processes are often modeled by assuming imaginal processes, highly task-specific assump-tions, or both. We propose the existence of a dedicated, unified cognitive system for the simulation of spatial processes,and show ways to model this system, including an ACT-R implementation that is currently in development. Results ofspatial cognition and brain-imaging research support this proposal. Operations of this system are proposed to be influencedby their complexity, which we assume to be a product of the extent and amount of necessary transformation steps. Thiscomplexity is further assumed to be limited in its extent, possibly explaining decision time effects between task difficul-ties in a mental folding task as being caused by cognitive re-encoding processes. A model for the mental folding tasklacking such a spatial system is presented, serving as a baseline to demonstrate the need of a system dedicated to mentaltransformations
- …
