arXiv subjects

Gian-Gabriel P. Garcia

Publications and source records attributed to Gian-Gabriel P. Garcia.

3 recordsLinked to original sources

arXiv2025

Neural Index Policies for Restless Multi-Action Bandits with Heterogeneous Budgets

Restless multi-armed bandits (RMABs) provide a scalable framework for sequential decision-making under uncertainty, but classical formulations assume binary actions and a single global budget. Real-world settings, such as healthcare, often involve multiple interventions with heterogeneous costs and constraints, where such assumptions break down. We introduce a Neural Index Policy (NIP) for multi-action RMABs with heterogeneous budget constraints. Our approach learns to assign budget-aware indices to arm--action pairs using a neural network, and converts them into feasible allocations via a differentiable knapsack layer formulated as an entropy-regularized optimal transport (OT) problem. The resulting model unifies index prediction and constrained optimization in a single end-to-end differentiable framework, enabling gradient-based training directly on decision quality. The network is optimized to align its induced occupancy measure with the theoretical upper bound from a linear programming relaxation, bridging asymptotic RMAB theory with practical learning. Empirically, NIP achieves near-optimal performance within 5% of the oracle occupancy-measure policy while strictly enforcing heterogeneous budgets and scaling to hundreds of arms. This work establishes a general, theoretically grounded, and scalable framework for learning index-based policies in complex resource-constrained environments.

cs.LG↗

arXiv2025

Tides Need STEMMED: A Locally Operating Spatio-Temporal Mutually Exciting Point Process with Dynamic Network for Improving Opioid Overdose Death Prediction

We develop a Spatio-TEMporal Mutually Exciting point process with Dynamic network (STEMMED), i.e., a point process network wherein each node models a unique community-drug event stream with a dynamic mutually-exciting structure, accounting for influences from other nodes. We show that STEMMED can be decomposed node-by-node, suggesting a tractable distributed learning procedure. Simulation shows that this learning algorithm can accurately recover known parameters of STEMMED, especially for small networks and long data-horizons. Next, we turn this node-by-node decomposition into an online cooperative multi-period forecasting framework, which is asymptotically robust to operational errors, to facilitate Opioid-related overdose death (OOD) trends forecasting among neighboring communities. In our numerical study, we parameterize STEMMED using individual-level OOD data and county-level demographics in Massachusetts. For any node, we observe that OODs within the same drug class from nearby locations have the greatest influence on future OOD trends. Furthermore, the expected proportion of OODs triggered by historical events varies greatly across counties, ranging between 30%-70%. Finally, in a practical online forecasting setting, STEMMED-based cooperative framework reduces prediction error by 60% on average, compared to well-established forecasting models. Leveraging the growing abundance of public health surveillance data, STEMMED can provide accurate forecasts of local OOD trends and highlight complex interactions between OODs across communities and drug types. Moreover, STEMMED enhances synergies between local and federal government entities, which is critical to designing impactful policy interventions.

stat.AP↗

arXiv2024

Who's the GOAT? Sports Rankings and Data-Driven Random Walks on the Symmetric Group

Given a collection of historical sports rankings, can one tell which player is the greatest of all time (i.e., the GOAT)? In this work, we design a data-driven random walk on the symmetric group to obtain a stationary distribution over player rankings, spanning across different time periods in sports history. We combine this distribution with a notion of stochastic dominance to obtain a partial order over the players. Compared to existing methods, our approach is distinct in that i) using historical rankings captures the evolution of value systems and facilitates player comparisons when head-to-head data is unavailable, and i) aggregating into a partial order formally comes to terms with the possibility that, while some player comparisons can be established conclusively, others are "too close to call." We implement our methods using publicly available data from the Association of Tennis Professionals (ATP) and the Women's Tennis Association (WTA). Our main findings indicate that Steffi Graf and Serena Williams are the ones that come ahead as the GOATs of the WTA. Likewise, the "Big Three," that is Novak Djokovic, Roger Federer, and Rafael Nadal, are the ones that come ahead as the GOATs of the ATP. As a secondary finding, we note major differences in terms of career and dominance longevity for top players across the associations. While initially motivated by this application in sports analytics, our methods can also be applied to other practical domains where deriving rankings from historical data can inform operational decisions, such as route planning logistics.

stat.AP↗