CGAWM - Curriculum-Guided Adversarial World Model for Data-Efficient Adversarial Reinforcement Learning
| dc.contributor.advisor | Beakal Gizachew | |
| dc.contributor.advisor | Natnael Argaw (Co-Advisor) | |
| dc.contributor.author | Migbar Abera Shibru | |
| dc.date.accessioned | 2026-08-26T14:54:01Z | |
| dc.date.available | 2026-08-26T14:54:01Z | |
| dc.date.issued | 2026-02 | |
| dc.description.abstract | Deep Reinforcement Learning (DRL) has achieved remarkable success in single-agent domains; however, Multi-Agent Reinforcement Learning (MARL) remains plagued by severe sample inefficiency and training instability, particularly in competitive zero-sum games. The non-stationarity inherent in adversarial environments where the opponent’s strategy evolves continuously often prevents agents from converging to optimal policies, leading to cycling behaviors or catastrophic forgetting. To address these challenges, this thesis proposes the Curriculum-Guided Adversarial World Model (CGAWM), a framework that integrates three methodological pillars: an Adversarial Dynamics Model (ADeM) that amplifies data efficiency via synthetic rollouts, an Adversarial Curriculum that stabilizes training by progressively increasing opponent competence, and an Uncertainty-Guided Exploration mechanism that incentivizes the agent to investigate high-entropy states. Evaluated on the PettingZoo simple adversary benchmark, experimental results demonstrate that CGAWM achieves a 136% improvement in asymptotic reward compared to state-of-the-art Model-Free PPO baselines. Furthermore, the proposed method exhibited superior reliability, achieving a consistentconvergence rate across random seeds where baseline methods failed to solve the task. Qualitative analysis confirmed the emergence of sophisticated game-theoretic behaviors, including Split-Coverage and Defensive Hovering, indicating that the agent successfully developed a primitive Theory of Mind. These findings suggest that integrating world models with structured curricula provides a scalable path toward robust autonomous systems in adversarial domains. | |
| dc.identifier.uri | https://etd.aau.edu.et/handle/123456789/8910 | |
| dc.language.iso | en | |
| dc.publisher | Addis Ababa University | |
| dc.subject | Multi-Agent Reinforcement Learning | |
| dc.subject | World Models | |
| dc.subject | Curriculum Learning | |
| dc.subject | Adversarial Training | |
| dc.subject | Sample Efficiency | |
| dc.title | CGAWM - Curriculum-Guided Adversarial World Model for Data-Efficient Adversarial Reinforcement Learning | |
| dc.type | Thesis |