Lucas Reymond

This project explores the use of Deep Q-Learning (DQN) to develop optimal epidemic mitigation strategies, minimizing deaths and infections during pandemics. The model is applied to simulate disease transmission across different cities in Switzerland, considering various interventions such as confinement, isolation, vaccination, and hospital capacity expansion.

The study compares the DQN-based policies against traditional strategies, such as Russo’s periodic confinement policy, demonstrating the superior effectiveness of the DQN approach in reducing mortality. The DQN agent learns dynamic policies that adapt to evolving pandemic scenarios, outperforming baseline policies by effectively balancing the duration of confinements and other interventions.

Findings

• Implemented binary action spaces (e.g. confinement vs. no confinement) and toggle-action spaces to manage complex state-action pairs efficiently.
• Enhanced policy learning using a decaying exploration rate to optimize cumulative rewards.
• Introduced a more realistic augmented action space, allowing simultaneous decisions on confinement, isolation, vaccination, and hospital bed expansion.
• Demonstrated that the DQN policy significantly reduced average deaths compared to the Russo policy, while maintaining optimal economic costs.
• Evaluated multiple performance metrics, including cumulative reward, mortality, and confinement duration, identifying the DQN policy as the most effective overall.

Impact

This project showcases the potential of Deep Reinforcement Learning for dynamic decision-making in epidemiological policy design, highlighting the effectiveness of adaptive strategies in mitigating pandemic impacts. The findings emphasize the importance of flexible, data-driven interventions in public health planning and disaster management.