| dc.description.abstract | Machine learning and artificial intelligence methods are commonly associated with robotics and gaming. In this dissertation, I investigated machine learning techniques in an entirely new field, disease outbreak management. In the first chapter, I explored current machine learning methods for disease outbreak management: regression, neural networks, dynamic programming, Monte Carlo methods, genetic algorithms, and network analysis. I discussed the strengths and weaknesses of each method, and proposed a flow diagram providing guidance to choose between the methods. In the second chapter, I proposed using deep Q-networks (DQN), which previously illustrated strong performance in larger decision spaces. For example, a previous study taught a learning agent to play the Atari 2600 games using only images of the game (Mnih et al., 2015). I modified the original DQN algorithm to address the objectives of disease management under resource constraints. I implemented DQN in four different case studies to explore management the 2001 United Kingdom foot and mouth disease (FMD) epidemic. Stability in the final policy was evident in the three case studies which investigated management with a smaller number of farms. The results of this chapter illustrated the limitations of using DQN for disease management when there were many farms to manage. A different computing architecture would be required to implement more training to improve performance in a larger decision setting. In the third chapter, I explored managing the 2001 FMD epidemic using simpler alternatives (e.g. - logistic regression, epidemiological interventions) compared to DQN. I concluded that DQN was clearly the optimal choice when there were a smaller number of farms, and simpler alternatives would be more feasible when there were more farms under consideration. Ultimately, the decision-maker would need to determine which method better met their management needs: superior performance from DQN at the cost of time and resources, or a faster and simpler method at the cost of suboptimal policies. | |
