{"id":{"repo_id":"gmu","oai_identifier":"oai:MARS:1920/14429"},"canonical_url":"https://search.dev.ndltd.org/etd/gmu/oai:MARS:1920/14429","repository":{"repo_id":"gmu","name":"George Mason University","base_url":"https://mars.gmu.edu/server/oai/request"},"display":{"title":"DECISION SUPPORT FOR DESIGN SYNTHESIS MODELED AS A PARTIALLY OBSERVED MARKOV DECISION PROCESS","abstract":"The purpose of this research is to improve the incorporation of risk management within the engineering design process. Current practices incorporate preliminary risk assessment but do not treat risk management holistically across system design and synthesis. Additionally, while most synthesis practices include pro forma verification activities, they do not provide guidance on the relative values of those activities, resulting in simultaneous generation of excess costs and false verification that the system meets requirements.The focus of this research is to demonstrate a proof-of-concept means of training decision support agents that will assist in optimally selecting design synthesis activities in a way that incorporates balanced risk avoidance, risk acceptance, and risk mitigation, including the costs (and value) of necessary verification activities. To that end, I have investigated the following hypotheses to test with this research: • Embedding the expected costs of system failures into a design’s state value function up-front will increase the likelihood of performing synthesis activities directed towards implementing risk mitigating features or of not performing synthesis activities directed towards implementing features that should be avoided. • Enforcing partial observability of the synthesis state when training decision support agents increases the relative value of performing verification activities. • Training a decision support agent against a design synthesis program will allow for one to determine, under various conditions, the relative preference of low-cost, low-accuracy testing to that of high-cost, high-accuracy testing. The proof-of-concept simulates synthesis of design specifications using a partially observed Markov decision process (POMDP). These simulations were then used to train decision support agents by means of reinforcement learning (RL) and approximate dynamic programming (ADP). After training, the agents were evaluated by recording their policy preferences given an observation of the design synthesis state. The results of the research demonstrated that agents trained in this way do decide to engage in risk avoidance activities as well as risk acceptance and mitigation, confirming the first hypothesis. However, there is not currently strong evidence to support the second or third hypotheses as the agents trained thus far rarely select to perform validation activities. The contributions of the research are as follows. This research provides a means to advance our understanding of the relative value of performing different verification activities at various points in the implementation of a feature in relation to sensitivity, specificity, and cost. And finally, this research advances the practice of risk management in design synthesis by demonstrating a method for optimally choosing between risk acceptance, mitigation, and avoidance during the design synthesis decision-making process.","abstract_html":"The purpose of this research is to improve the incorporation of risk management within the engineering design process. Current practices incorporate preliminary risk assessment but do not treat risk management holistically across system design and synthesis. Additionally, while most synthesis practices include pro forma verification activities, they do not provide guidance on the relative values of those activities, resulting in simultaneous generation of excess costs and false verification that the system meets requirements.The focus of this research is to demonstrate a proof-of-concept means of training decision support agents that will assist in optimally selecting design synthesis activities in a way that incorporates balanced risk avoidance, risk acceptance, and risk mitigation, including the costs (and value) of necessary verification activities. To that end, I have investigated the following hypotheses to test with this research: • Embedding the expected costs of system failures into a design’s state value function up-front will increase the likelihood of performing synthesis activities directed towards implementing risk mitigating features or of not performing synthesis activities directed towards implementing features that should be avoided. • Enforcing partial observability of the synthesis state when training decision support agents increases the relative value of performing verification activities. • Training a decision support agent against a design synthesis program will allow for one to determine, under various conditions, the relative preference of low-cost, low-accuracy testing to that of high-cost, high-accuracy testing. The proof-of-concept simulates synthesis of design specifications using a partially observed Markov decision process (POMDP). These simulations were then used to train decision support agents by means of reinforcement learning (RL) and approximate dynamic programming (ADP). After training, the agents were evaluated by recording their policy preferences given an observation of the design synthesis state. The results of the research demonstrated that agents trained in this way do decide to engage in risk avoidance activities as well as risk acceptance and mitigation, confirming the first hypothesis. However, there is not currently strong evidence to support the second or third hypotheses as the agents trained thus far rarely select to perform validation activities. The contributions of the research are as follows. This research provides a means to advance our understanding of the relative value of performing different verification activities at various points in the implementation of a feature in relation to sensitivity, specificity, and cost. And finally, this research advances the practice of risk management in design synthesis by demonstrating a method for optimally choosing between risk acceptance, mitigation, and avoidance during the design synthesis decision-making process.","abstract_has_math":false,"creators":["Florer, Luke"],"institution":null,"degree_name":null,"degree_level":null,"degree_discipline":null,"degree_department":null,"school":null,"contributors":[],"advisors":[],"committee_chairs":[],"committee_members":[],"year":2024,"date_issued":"2024","date_published":"2024","updated_at":"2026-07-27T19:52:10Z","subjects":["decision support","deep reinforcement learning","engineering design","partially observed Markov decision process"],"languages":[],"rights":[],"rights_urls":[],"identifier_entries":[{"key":"dc:identifier","label":"Identifier","values":["hdl:1920/14429"],"render_values":[{"text":"hdl:1920/14429","href":null,"code":true}]}]},"links":{"outbound_url":null,"outbound_label":null,"outbound_source":null},"metadata_groups":[{"id":"academic_context","label":"Academic Context","entries":[{"key":"dc:date.issued","label":"Date","values":["2024"]},{"key":"dc:type","label":"Dc Type","values":["Dissertation"]}]},{"id":"subjects_keywords","label":"Subjects and Keywords","entries":[{"key":"dc:subject","label":"Dc Subject","values":["decision support","deep reinforcement learning","engineering design","partially observed Markov decision process"]}]},{"id":"identifiers","label":"Identifiers","entries":[{"key":"dc:identifier","label":"Identifier","values":["hdl:1920/14429"]}]},{"id":"additional","label":"Additional Metadata","entries":[{"key":"dc:description.other","label":"Dc Description Other","values":["The purpose of this research is to improve the incorporation of risk management within the engineering design process. Current practices incorporate preliminary risk assessment but do not treat risk management holistically across system design and synthesis. Additionally, while most synthesis practices include pro forma verification activities, they do not provide guidance on the relative values of those activities, resulting in simultaneous generation of excess costs and false verification that the system meets requirements.The focus of this research is to demonstrate a proof-of-concept means of training decision support agents that will assist in optimally selecting design synthesis activities in a way that incorporates balanced risk avoidance, risk acceptance, and risk mitigation, including the costs (and value) of necessary verification activities. To that end, I have investigated the following hypotheses to test with this research: • Embedding the expected costs of system failures into a design’s state value function up-front will increase the likelihood of performing synthesis activities directed towards implementing risk mitigating features or of not performing synthesis activities directed towards implementing features that should be avoided. • Enforcing partial observability of the synthesis state when training decision support agents increases the relative value of performing verification activities. • Training a decision support agent against a design synthesis program will allow for one to determine, under various conditions, the relative preference of low-cost, low-accuracy testing to that of high-cost, high-accuracy testing. The proof-of-concept simulates synthesis of design specifications using a partially observed Markov decision process (POMDP). These simulations were then used to train decision support agents by means of reinforcement learning (RL) and approximate dynamic programming (ADP). After training, the agents were evaluated by recording their policy preferences given an observation of the design synthesis state. The results of the research demonstrated that agents trained in this way do decide to engage in risk avoidance activities as well as risk acceptance and mitigation, confirming the first hypothesis. However, there is not currently strong evidence to support the second or third hypotheses as the agents trained thus far rarely select to perform validation activities. The contributions of the research are as follows. This research provides a means to advance our understanding of the relative value of performing different verification activities at various points in the implementation of a feature in relation to sensitivity, specificity, and cost. And finally, this research advances the practice of risk management in design synthesis by demonstrating a method for optimally choosing between risk acceptance, mitigation, and avoidance during the design synthesis decision-making process."]},{"key":"dc:title","label":"Title","values":["DECISION SUPPORT FOR DESIGN SYNTHESIS MODELED AS A PARTIALLY OBSERVED MARKOV DECISION PROCESS"]}]}],"canonical_facts":{"dc:date.issued":["2024"],"dc:description.other":["The purpose of this research is to improve the incorporation of risk management within the engineering design process. Current practices incorporate preliminary risk assessment but do not treat risk management holistically across system design and synthesis. Additionally, while most synthesis practices include pro forma verification activities, they do not provide guidance on the relative values of those activities, resulting in simultaneous generation of excess costs and false verification that the system meets requirements.The focus of this research is to demonstrate a proof-of-concept means of training decision support agents that will assist in optimally selecting design synthesis activities in a way that incorporates balanced risk avoidance, risk acceptance, and risk mitigation, including the costs (and value) of necessary verification activities. To that end, I have investigated the following hypotheses to test with this research: • Embedding the expected costs of system failures into a design’s state value function up-front will increase the likelihood of performing synthesis activities directed towards implementing risk mitigating features or of not performing synthesis activities directed towards implementing features that should be avoided. • Enforcing partial observability of the synthesis state when training decision support agents increases the relative value of performing verification activities. • Training a decision support agent against a design synthesis program will allow for one to determine, under various conditions, the relative preference of low-cost, low-accuracy testing to that of high-cost, high-accuracy testing. The proof-of-concept simulates synthesis of design specifications using a partially observed Markov decision process (POMDP). These simulations were then used to train decision support agents by means of reinforcement learning (RL) and approximate dynamic programming (ADP). After training, the agents were evaluated by recording their policy preferences given an observation of the design synthesis state. The results of the research demonstrated that agents trained in this way do decide to engage in risk avoidance activities as well as risk acceptance and mitigation, confirming the first hypothesis. However, there is not currently strong evidence to support the second or third hypotheses as the agents trained thus far rarely select to perform validation activities. The contributions of the research are as follows. This research provides a means to advance our understanding of the relative value of performing different verification activities at various points in the implementation of a feature in relation to sensitivity, specificity, and cost. And finally, this research advances the practice of risk management in design synthesis by demonstrating a method for optimally choosing between risk acceptance, mitigation, and avoidance during the design synthesis decision-making process."],"dc:identifier":["hdl:1920/14429"],"dc:subject":["decision support","deep reinforcement learning","engineering design","partially observed Markov decision process"],"dc:title":["DECISION SUPPORT FOR DESIGN SYNTHESIS MODELED AS A PARTIALLY OBSERVED MARKOV DECISION PROCESS"],"dc:type":["Dissertation"]},"updated_at":"2026-07-27T19:52:10Z"}