Author ORCID Identifier
https://orcid.org/0000-0002-2578-5137
Document Type
Theses, Ph.D
Disciplines
Computer Sciences
Abstract
Safety-critical systems, which are crucial for human safety and the environment, are difficult to control and operate. Traditional controllers need precise models of these complex systems, which is hard to develop. \acrfull{drl} offers a potential solution by learning from interactions rather than detailed models, but it faces limitations such as non-transparent decision-making and an expensive, unsafe learning process. Additionally, a key challenge in DRL is ensuring effective decision-making in rare situations.
This thesis proposes a novel approach called the \acrfull{prop_frame} that enables safe and reliable control of critical systems. SRLA combines probabilistic modelling with reinforcement learning to create an interpretable system that can focus on the filtered state space. SRLA is activated in specific situations identified autonomously through the combination of probabilistic modelling and DRL, such as when the system is in an abnormal state or performing a sub-task. It uses policy cloning to initialise a baseline policy, which minimises the need for expensive exploration. Additionally, SRLA works alongside conventional control strategies to ensure safe and reliable decision-making.
SRLA's effectiveness is demonstrated through diverse safety-critical industrial case studies. It outperforms other methods in the predictive maintenance of turbofan engines by accurately predicting failures and identifying health states and root causes. In process control, SRLA autonomously synchronizes with conventional controllers and activates in critical situations. As a control room decision support for human operators, SRLA provides real-time suggestions to help operators avoid failures and can predict human errors using process data and human-computer interaction logs. In the case of industrial robotics, SRLA enables robots to learn complex tasks by breaking them into specialised sub-tasks in simulation and safely transferring the policies to the real world, outperforming traditional DRL.
DOI
https://doi.org/10.21427/p05p-az54
Recommended Citation
Abbas, Ammar, "A Hierarchical Framework for Interpretable, Safe, and Specialised Deep Reinforcement Learning" (2024). Doctoral. 17.
https://arrow.tudublin.ie/ittthedoc/17
Funder
Marie Skłodowska-Curie International Training Network European Industrial Doctorate (MSCA-ITN-EID)
Creative Commons License
This work is licensed under a Creative Commons Attribution-NonCommercial-Share Alike 4.0 International License.
Publication Details
A dissertation submitted in fulfilment of the requirements of Technological University Dublin for the Ph.D. in Artificial Intelligence (Deep Reinforcement Learning). Technological University Dublin, 2024.
doi:10.21427/p05p-az54