Showing papers by "Brian Veitch published in 2019"

Situation awareness modeling for emergency management on offshore platforms

Abstract: Situation awareness is the first and most important step in emergency management. It is a dynamic step involving evolving conditions and environments. It is an area of active research. This study presents a Markov Logic Network to model SA focusing on fire accidents and emergency evacuation. The model has been trained using empirical data obtained from case studies. The case studies involved human participants who were trained for responding to emergencies involving fire and smoke using a virtual environment. The simulated (queried) and empirical findings are reasonably consistent. The proposed model enables implementing an agent that exploits environmental cues and cognitive states to determine the type of emergency currently being faced. Considering each emergency type as a situation, the model can be used to develop a repertoire of situations for agents so that the repertoire can act as an agent’s experience for later decision-making.

Human-Like Sequential Learning of Escape Routes for Virtual Reality Agents

Abstract: The Piper Alpha disaster (1988) witnessed 167 casualties. The offshore safety guidelines developed afterward highlighted the need for effective and regular training to overcome the problems in evacuation procedures. Today, virtual environments are effective training platforms due to high-end audio/visual and interactive capabilities. These virtual environments exploit agents with human-like steering capabilities, but with limited or no capacity to learn routes. This work proposes a sequential route learning methodology for agents that resembles the way people learn routes. The methodology developed here exploits a generalized stochastic Petri-net based route learning model iteratively. The simulated results are compared with the route learning strategies of human participants. The data on human participants were collected by the authors from an earlier study in a virtual environment. The main contribution lies in modeling people’s route learning behavior over the course of successive exposures. It is found that the proposed methodology models human-like sequential route learning if there are no easy detours from the original escape route. Although the model does not accurately capture individual learning strategies for all decision nodes, it can be used as a model of compliant, rule-following training guides for a virtual environment.

On the realization of the recognition-primed decision model for artificial agents

Abstract: This work proposes a methodology to program an artificial agent that can make decisions based on a naturalistic decision-making approach called recognition-primed decision model (RPDM). The proposed methodology represents the main constructs of RPDM in the language of Belief-Desire-Intention logic. RPDM considers decision-making as a synthesis of three phenomenal abilities of the human mind. The first is one’s use of experience to recognize a situation and suggest appropriate responses. The main concern here is on situation awareness because the decision-maker needs to establish that a current situation is the same or similar to one previously experienced, and the same solution is likely to work this time too. To this end, the proposed modeling approach uses a Markov logic network to develop an Experiential-Learning and Decision-Support module. The second component of RPDM deals with the cases when a decision-maker’s experience becomes secondary because the situation has not been recognized as typical. In this case, RPDM suggests a diagnostic mechanism that involves feature-matching, and, therefore, an ontology (of the domain of interest) based reasoning approach is proposed here to deal with all such cases. The third component of RPDM is the proposal that human beings use intuition and imagination (mental stimulation) to make sure whether a course of action should work in a given situation or not. Mental simulation is modeled here as a Bayesian network that computes the probability of occurrence of an effect when a cause is more likely. The agent-based model of RPDM has been validated with real (empirical) data to compare the simulated and empirical results and develop a correspondence in terms of the value of the result, as well as the reasoning.

Using Simulator Data to Facilitate Human Reliability Analysis

Abstract: Data scarcity has always been a significant challenge in the domain of human reliability analysis (HRA). The advancement of simulation technologies provides opportunities to collect human performance data that can facilitate both the development and validation paradigms of HRA. The potential of simulator data to improve HRA can be tapped through the use of advanced machine learning tools like Bayesian methods. Except for Bayesian networks, Bayesian methods have not been widely used in the HRA community. This paper uses a Bayesian method to enhance human error probability (HEP) assessment in offshore emergency situations using data generated in a simulator. Assessment begins by using constrained noninformative priors to define the HEPs in emergency situations. An experiment is then conducted in a simulator to collect human performance data in a set of emergency scenarios. Data collected during the experiment are used to update the priors and obtain informed posteriors. Use of the informed posteriors enables better understanding of the performance, and a more reliable and objective assessment of human reliability, compared to traditional assessment using expert judgment.

Validating Human Behavior Representation Model of General Personnel During Offshore Emergency Situations

Abstract: With the advancement of simulation-based training, intelligent agents that can display human-like behavior have become common. From military combat simulations to nuclear power plant simulation, agents have been widely used to facilitate team training (as team mates, opponents, or both). Credibility of these agents is vital to ensure a sound training process. Credibility of the agents largely depends on the credibility of the underlying human behavior representation (HBR) model. This is why validation of the HBR model is necessary to ensure realistic agent behavior. However, the non-deterministic nature of the HBR and the subjectivity in experts’ judgment during the validation process make HBR model validation more challenging compared to physics based models. This paper presents the validation process of an HBR model of general personnel created for use in an offshore emergency training simulator. Three types of agents (naive, in-between, and ideal) are created in the simulator using the HBR model. The paper discusses the use of empirical evidence as referents, along with subject matter experts. A two-level three factor experiment was conducted using 36 participants. Several performance metrics were collected during the experiment, including route selected for evacuation, time to muster, time spent running, interaction with fire doors and watertight doors, interaction with hazards, and reporting to the muster station. Data collected during the experimental study have been used in this paper to demonstrate how the use of empirical evidence can facilitate HBR validation. High-level tasks performed during HBR validation are discussed in detail. Special emphasis is given on acceptability criteria testing to ensure that the HBR model performs adequately under different operating conditions. Results show that the proposed HBR model meets the acceptability criteria requirement for all types of agents. In general, the ideal agents exhibited safe behavior during offshore emergency egress, whereas the naive and in-between agents showed erroneous behavior at times. For example, during the simulation runs of a critical emergency scenario where the primary egress route was obstructed by a hazard, the ideal agents either waited and listened to the public address announcement and followed an alternative egress route (60% cases), or they initially chose their preferred route but re-routed immediately after encountering the hazard (40% cases). In all cases, the in-between agents started with their preferred route and re-routed after encountering the hazard, and the naive agents proceeded with their preferred route even when the route was compromised.

Assessment of changes in neural activity during acquisition of spatial knowledge using EEG signal classification.

Abstract: Objective. This study explored the classification of electroencephalography (EEG) signals to assess changes in neural activity as individuals performed a training task in a virtual environment simulator. Commonly, task behavior and perception are used to assess a trainee's ability to perform a task, however, changes in cognition are not usually measured and could be important to provide a true indication of an individual's level of knowledge or skill. Approach. In this study, 15 participants acquired spatial knowledge via 60 navigation trials (divided into ten blocks) in a novel virtual environment. Time performance, perceived certainty, and EEG signal data were collected. Main results. A significant increase in alpha power and classification accuracy of EEG data from block 1 against blocks 2–10 was observed and stabilized after block 7, while time performance and perceived certainty measures improved and stabilized after block 5 and 6, respectively. Significance. Results suggest that changes in neural activity, which may reflect an increase in cognitive efficiency, could provide additional insight beyond time performance and perceived certainty.

Being prepared for emergencies: a virtual environment experiment on the retention and maintenance of egress skills

Abstract: The retention of safety-critical egress skills is an essential part of emergency preparedness on offshore petroleum platforms. Virtual environment (VE) training has been shown to be an effective method for teaching basic onboard familiarization and offshore emergency evacuation procedures. This technology has the potential to train crews before they are deployed offshore. This paper investigates the long-term retention and maintenance of emergency egress competence obtained using a virtual offshore platform. In particular, the research aimed to answer two questions: (1) what egress skills can be remembered after a period of 6 months? and (2) how effective is a VE-based retraining program at maintaining egress skills? A two-phased experiment was designed to first teach basic egress skills and subsequently assess skill retention after a 6- to 9-month period. The first phase of the experiment used a simulation-based mastery learning (SBML) pedagogical approach to teach naive subjects the necessary spatial and procedural skills to evacuate safely. In the second phase of the experiment, the same 36 participants were tested after the retention interval on their ability to respond to a series of egress test scenarios. Participants who had trouble remembering the egress procedures were provided retraining on deficient skills. The results of the experiment indicate that emergency egress skills (both spatial and procedural knowledge) are susceptible to skill decay. This paper will highlight the skills that were most susceptible to skill fade after a period of 6 to 9 months and discuss the efficacy of the retraining participants received to return to competence.