Analysis of deep reinforcement learning for resilient autonomous space systems
Loading...
Date
2026-06-22
Authors
Λαχανάς, Άγγελος
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
This diploma thesis examines the interpretability of deep reinforcement learning agents in the context of autonomous space systems. As autonomous systems are increasingly introduced into high-criticality applications, their evaluation cannot rely only on external performance indicators. From the perspective of resilience engineering, it is important to examine not only whether a system completes its mission successfully, but also how its internal representation changes under nominal, disturbed, and degraded operating conditions. In this work, a deep reinforcement learning agent is trained in a simulated autonomous landing environment. The latent representations produced by the trained policy during multiple rollout episodes are then analyzed. The methodology combines linear dimensionality-reduction techniques, nonlinear manifold-learning methods, and clustering in order to investigate the geometric, statistical, and behavioral structure of the policy latent space. The resulting projections and clusters are compared with rollout variables such as system state, selected actions, episode time, reward, and mission outcome. The central objective of the thesis is to investigate whether the internal representations of a trained agent can be used as a diagnostic tool for analyzing the resilience
of an autonomous system. For this purpose, the nominal latent structure is used as a reference and compared with the agent’s response under disturbances and degradation mechanisms. This makes it possible to examine whether reductions in external performance, measured through return and success rate, are accompanied by measurable changes in latent-space geometry, behavioral-mode occupancy, and transitions between latent modes. The thesis contributes to the connection between neural-policy interpretability and resilience engineering for autonomous systems. Instead of treating the neural network purely as a black box, the proposed analysis uses the policy latent space to examine the internal organization of the agent’s behavior, thus providing additional insight into how a trained agent maintains, adapts, or loses its ability to complete the mission under disturbed operating conditions.
Description
Keywords
Βαθιά ενισχυτική μάθηση, Μάθηση πολλαπλοτήτων, Λανθάνουσες αναπαραστάσεις, Αυτόνομα συστήματα, Ανθεκτικά συστήματα, Συστήματα υψηλής κρισιμότητας