Research Interests

I have developed models to understand and control diffusion dynamics. In one line of work, I created stochastic and mean-field models, which describe the average behavior of large systems, to analyze competitive product adoption when local network effects matter. These models explain how rival products spread across multiple, interrelated networks and how the correlation between networks influences long-term market shares and optimal pricing. This framework not only clarifies when competing products can coexist but also guides firms in designing pricing and advertising policies that exploit or counteract network externalities to maximize long-run profits. In a complementary project, I examined profit-maximizing viral marketing campaigns where incentivized consumers act as information spreaders. By combining epidemic-style approximations with nonlinear optimization, I identified cost-effective incentive policies and showed how network topology and spectral properties govern the speed and reach of information cascades.

To actively control these diffusion processes, I design reinforcement learning frameworks that treat the evolving network itself as a system to observe and influence. In recent work on dynamic pricing and advertising under network effects, I first built a mathematical model, similar in spirit to models of disease spread, that describes how consumers move between stages of interest and adoption. From this, I derived a simplified "average-behavior" representation to identify when a product is likely to take off. Building on this foundation, I developed a new RL approach that merges a learning-by-trial-and-error framework (actor–critic RL) with a graph-based encoder, which automatically summarizes complex social network connections so the RL system can make decisions without being overwhelmed by detail. Using a high-performance GPU-based simulator, this network-aware controller learns to adaptively choose pricing and advertising actions and consistently outperforms approaches that ignore network structure, especially in networks with uneven patterns of connectivity. This work shows how reinforcement learning and graph-based modeling can be combined to create scalable, adaptive AI methods that both capture the structure of social networks and use that understanding to make better real-time decisions.

I have also extended these ideas to situations where the goal is to contain harmful diffusion, such as stopping the spread of infectious diseases. I built a stochastic network-based model of disease spread that reflects key realities such as breakthrough infections and waning immunity and used it to identify the conditions under which a disease will either persist or die out. On top of this model, I created a new reinforcement learning algorithm that learns how to allocate vaccines in real time. Both the "decision maker" and the "evaluator" within the algorithm are equipped with graph-based neural networks, which automatically learn how people are connected and how those connections evolve as the epidemic unfolds. A custom GPU-accelerated simulator allows the system to safely experiment and adapt its strategy. The resulting algorithm learns to target vaccinations to the most influential individuals at the right time, even when only partial information about the network is available, and consistently outperforms standard heuristics such as vaccinating at random or simply choosing the most connected people. This research demonstrates how reinforcement learning and graph neural networks can be combined to create practical AI methods that not only model the spread of disease but also discover effective ways to control real-world contagion.

I have integrated machine learning with classical optimization to solve large-scale operations problems. In work on manual order-picking operations, I introduced a hybrid algorithm that combines a neural network, a genetic algorithm, and a classical column generation technique to minimize the total time needed to complete warehouse picking. The key innovation is using a neural network to estimate the optimal travel distance for pickers, which dramatically reduces the computational bottleneck that traditionally limits column generation methods. The neural estimator feeds into a genetic algorithm that searches for new solutions with lower cost, enabling the overall method to handle real-world instances that defeat exact mixed-integer programming solvers. This project illustrates how AI models can be blended with established operations-research techniques to create algorithms that are both computationally efficient and theoretically sound.