{"id":{"repo_id":"cape-town","oai_identifier":"oai:open.uct.ac.za:11427/41574"},"canonical_url":"https://search.dev.ndltd.org/etd/cape-town/oai:open.uct.ac.za:11427/41574","repository":{"repo_id":"cape-town","name":"University of Cape Town","base_url":"https://open.uct.ac.za/oai/request"},"display":{"title":"Self-attention policy architectures for reinforcement learning under partial observability","abstract":"Intermittent unavailability of sensory signals due to sensor failure and/or latency is a problem encountered in production environments such as in large manufacturing plants, for example. Deep reinforcement learning offers a natural solution for process control and optimisation in such environments. However, a shortcom-ing of conventional agent policy architectures in this instance is an inability to handle variable-sized inputs composed of available sensory signals, thus requiring the imputation of unavailable sensory signals with data which necessarily constitutes noise. We explore self-attention-based policy architectures as a solution to this problem, demonstrating their robustness under conditions of high partial observability on different rein-forcement learning benchmark tasks, and explore the advantages and disadvantages offered by our solution over conventional policy architectures. Additionally, we propose a novel hard attention mechanism, used in conjunction with our proposed policy architecture, enabling the agent to attend to the most salient sensory signals and allowing for greater interpretability of the agent's decision-making.","abstract_html":"Intermittent unavailability of sensory signals due to sensor failure and/or latency is a problem encountered in production environments such as in large manufacturing plants, for example. Deep reinforcement learning offers a natural solution for process control and optimisation in such environments. However, a shortcom-ing of conventional agent policy architectures in this instance is an inability to handle variable-sized inputs composed of available sensory signals, thus requiring the imputation of unavailable sensory signals with data which necessarily constitutes noise. We explore self-attention-based policy architectures as a solution to this problem, demonstrating their robustness under conditions of high partial observability on different rein-forcement learning benchmark tasks, and explore the advantages and disadvantages offered by our solution over conventional policy architectures. Additionally, we propose a novel hard attention mechanism, used in conjunction with our proposed policy architecture, enabling the agent to attend to the most salient sensory signals and allowing for greater interpretability of the agent's decision-making.","abstract_has_math":false,"creators":["Du Plessis, Jeremy"],"institution":"Department of Mathematics and Applied Mathematics","degree_name":null,"degree_level":null,"degree_discipline":null,"degree_department":null,"school":null,"contributors":[],"advisors":["Shock, Jonathan"],"committee_chairs":[],"committee_members":[],"year":2025,"date_issued":"2025","date_published":"2025","updated_at":"2026-07-24T01:34:01Z","subjects":["Self-attention"],"languages":["en"],"rights":[],"rights_urls":[],"identifier_entries":[]},"links":{"outbound_url":"http://hdl.handle.net/11427/41574","outbound_label":"Handle","outbound_source":"dc:identifier.uri"},"metadata_groups":[{"id":"people","label":"People","entries":[{"key":"dc:contributor.advisor","label":"Advisor","values":["Shock, Jonathan"]},{"key":"dc:creator","label":"Author","values":["Du Plessis, Jeremy"]}]},{"id":"academic_context","label":"Academic Context","entries":[{"key":"dc:date.accessioned","label":"Dc Date Accessioned","values":["2025-08-13T12:59:08Z"]},{"key":"dc:date.available","label":"Dc Date Available","values":["2025-08-13T12:59:08Z"]},{"key":"dc:date.issued","label":"Date","values":["2025"]},{"key":"dc:publisher.department","label":"Dc Publisher Department","values":["Department of Mathematics and Applied Mathematics"]},{"key":"dc:publisher.institution","label":"Dc Publisher Institution","values":["University of Cape Town"]},{"key":"dc:type","label":"Dc Type","values":["Thesis / Dissertation"]},{"key":"dc:type.qualificationlevel","label":"Dc Type Qualificationlevel","values":["Masters","MSc"]}]},{"id":"subjects_keywords","label":"Subjects and Keywords","entries":[{"key":"dc:subject","label":"Dc Subject","values":["Self-attention"]}]},{"id":"language_rights","label":"Language and Rights","entries":[{"key":"dc:language.iso","label":"Language (ISO)","values":["en"]}]},{"id":"identifiers","label":"Identifiers","entries":[{"key":"dc:identifier.uri","label":"Identifier URI","values":["http://hdl.handle.net/11427/41574"]}]},{"id":"additional","label":"Additional Metadata","entries":[{"key":"dc:description.abstract","label":"Abstract","values":["Intermittent unavailability of sensory signals due to sensor failure and/or latency is a problem encountered in production environments such as in large manufacturing plants, for example. Deep reinforcement learning offers a natural solution for process control and optimisation in such environments. However, a shortcom-ing of conventional agent policy architectures in this instance is an inability to handle variable-sized inputs composed of available sensory signals, thus requiring the imputation of unavailable sensory signals with data which necessarily constitutes noise. We explore self-attention-based policy architectures as a solution to this problem, demonstrating their robustness under conditions of high partial observability on different rein-forcement learning benchmark tasks, and explore the advantages and disadvantages offered by our solution over conventional policy architectures. Additionally, we propose a novel hard attention mechanism, used in conjunction with our proposed policy architecture, enabling the agent to attend to the most salient sensory signals and allowing for greater interpretability of the agent's decision-making."]},{"key":"dc:title","label":"Title","values":["Self-attention policy architectures for reinforcement learning under partial observability"]}]}],"canonical_facts":{"dc:contributor.advisor":["Shock, Jonathan"],"dc:creator":["Du Plessis, Jeremy"],"dc:date.accessioned":["2025-08-13T12:59:08Z"],"dc:date.available":["2025-08-13T12:59:08Z"],"dc:date.issued":["2025"],"dc:description.abstract":["Intermittent unavailability of sensory signals due to sensor failure and/or latency is a problem encountered in production environments such as in large manufacturing plants, for example. Deep reinforcement learning offers a natural solution for process control and optimisation in such environments. However, a shortcom-ing of conventional agent policy architectures in this instance is an inability to handle variable-sized inputs composed of available sensory signals, thus requiring the imputation of unavailable sensory signals with data which necessarily constitutes noise. We explore self-attention-based policy architectures as a solution to this problem, demonstrating their robustness under conditions of high partial observability on different rein-forcement learning benchmark tasks, and explore the advantages and disadvantages offered by our solution over conventional policy architectures. Additionally, we propose a novel hard attention mechanism, used in conjunction with our proposed policy architecture, enabling the agent to attend to the most salient sensory signals and allowing for greater interpretability of the agent's decision-making."],"dc:identifier.uri":["http://hdl.handle.net/11427/41574"],"dc:language.iso":["en"],"dc:publisher.department":["Department of Mathematics and Applied Mathematics"],"dc:publisher.institution":["University of Cape Town"],"dc:subject":["Self-attention"],"dc:title":["Self-attention policy architectures for reinforcement learning under partial observability"],"dc:type":["Thesis / Dissertation"],"dc:type.qualificationlevel":["Masters","MSc"]},"updated_at":"2026-07-24T01:34:01Z"}