02 / Research topics
Potential PhD Research Topics in Optimization
Explore suggested topics and the methods, tools and expertise that could support each investigation.
1
Scalable Bilevel Optimization for Learning Decisions Under Strategic and Distributional Uncertainty
Modern learning-enabled decisions often involve nested optimization while data distributions and stakeholders change strategically. Existing bilevel methods remain computationally expensive, statistically fragile, and weakly justified outside ideal assumptions. This project could develop scalable algorithms using implicit differentiation, variance reduction, robust optimization, and adaptive sampling. Objectives include convergence guarantees, sensitivity analysis, strategic robustness, and efficient implementations for large datasets. Evaluation should compare solution quality, computational cost, and stability across hyperparameter learning, market design, and resource allocation benchmarks.
Applications and use cases
2
Trustworthy Federated Optimization with Privacy, Fairness, and Heterogeneous Client Participation
Federated systems must optimize shared models despite nonidentical data, unreliable participation, communication limits, and adversarial clients. Current approaches often trade accuracy for privacy or fairness and provide limited guarantees under realistic heterogeneity. The proposed research should design adaptive aggregation and client-selection methods with formal privacy, convergence, robustness, and fairness guarantees. Experiments can examine mobile, healthcare, and cross-organizational datasets under communication failures and poisoning attacks. The contribution should unite optimization theory with measurable governance requirements rather than treating privacy and fairness as afterthoughts.
Applications and use cases
3
Differentiable Combinatorial Optimization for Real Time Decisions in Dynamic Operational Systems
Many operational decisions are combinatorial, yet modern predictive models require gradients to learn from downstream decisions. Existing differentiable relaxations can be biased, unstable, or computationally impractical for changing constraints. This research should develop principled differentiable surrogates, implicit layers, or hybrid exact-learning algorithms that preserve discrete feasibility and improve end-to-end decisions. Objectives include approximation bounds, efficient training, robustness to prediction error, and real-time deployment. Suitable evaluations include routing, scheduling, matching, inventory, and network design with simulated and industry-inspired workloads.
Applications and use cases
4
Safe Reinforcement Learning Through Constrained and Distributionally Robust Optimization
Reinforcement learning can discover effective policies but remains unsafe when exploration violates physical, financial, or societal constraints. Existing safe methods commonly assume known dynamics, abundant data, or reliable constraint estimates. This project should combine constrained Markov decision processes with distributionally robust and risk-sensitive optimization to produce policies with finite-sample safety guarantees. Key objectives are sample efficiency, adaptation under model uncertainty, interpretable risk controls, and performance retention. Testing should use simulators and real-world-inspired environments in robotics, energy, healthcare, and autonomous systems.
Applications and use cases
5
Optimization of Foundation Models for Efficient Adaptation, Routing, and Inference
Foundation models deliver broad capabilities but demand substantial computation, memory, and energy for adaptation and inference. Current optimization methods often improve one metric while worsening latency, accuracy, robustness, or environmental cost. This proposal could formulate model adaptation, expert routing, quantization, pruning, and serving as joint constrained optimization problems. Objectives include theoretically informed algorithms, hardware-aware guarantees, dynamic resource allocation, and transparent quality-efficiency tradeoffs. Validation should span language and multimodal models, deployment platforms, and workload distributions rather than relying solely on benchmark accuracy.
Applications and use cases
6
Distributionally Robust Optimization for Climate Resilient Infrastructure and Energy Planning
Climate volatility makes historical data unreliable for infrastructure investment and energy operation. Conventional stochastic models frequently underestimate tail events and cannot represent ambiguity about future distributions. The research should develop tractable distributionally robust formulations using climate projections, spatial dependence, and infrastructure network structure. Novelty may come from ambiguity-set learning, decomposition algorithms, and decision-dependent uncertainty. Objectives include out-of-sample reliability, interpretable risk budgets, scalability, and equitable resilience. Case studies could address power systems, water networks, transportation, and disaster response under extreme scenarios.
Applications and use cases
7
Quantum and Quantum Inspired Optimization for Large Scale Discrete Decision Problems
Quantum optimization attracts attention for potentially improving difficult discrete decisions, but practical advantage remains uncertain and problem embeddings can dominate performance. Many studies lack fair classical baselines, noise-aware analysis, or realistic scaling evidence. This PhD should investigate hybrid quantum-classical algorithms and quantum-inspired heuristics for structured combinatorial problems, emphasizing rigorous benchmarking and complexity-aware formulation. Objectives include identifying favorable problem classes, improving decomposition and parameter tuning, quantifying noise effects, and establishing reproducible comparisons across current hardware and high-quality classical solvers.
Applications and use cases
8
Causal Optimization for Personalized Policies Under Interventions and Longitudinal Data
Optimization based on correlations can recommend actions that fail when interventions alter outcomes or populations. Causal decision methods remain limited by hidden confounding, treatment constraints, time dependence, and imperfect experimentation. This research should integrate causal inference with robust and sequential optimization to learn policies that maximize outcomes under realistic interventions. Objectives include identification conditions, uncertainty-aware policy optimization, fairness analysis, and sample-efficient offline learning. Applications may use healthcare, education, labor, or public policy data, with sensitivity analyses that make assumptions explicit.
Applications and use cases
9
Optimization Algorithms for Decentralized Multi Agent Systems with Communication and Privacy Constraints
Decentralized agents must coordinate without a central authority while facing delayed communication, unreliable networks, privacy requirements, and local objectives. Existing consensus optimization methods often assume synchronous operation and benign participants. This project should develop asynchronous, privacy-preserving, and resilient algorithms for nonconvex and constrained multi-agent optimization. Desired contributions include convergence and communication guarantees, adaptive topology management, and deployment-ready protocols. Experiments should use robot teams, sensor networks, distributed learning, and market simulations to measure efficiency, resilience, privacy leakage, and coordination quality.
Applications and use cases
10
Fair and Explainable Optimization for Resource Allocation in High Stakes Socio Technical Systems
Optimization systems increasingly allocate loans, healthcare, jobs, housing, and public resources, yet objectives and constraints can reproduce inequity. Existing fairness formulations may conflict with efficiency, obscure tradeoffs, or ignore structural disadvantage and feedback effects. This research should develop multi-objective and participatory optimization methods with auditable explanations and context-sensitive fairness guarantees. Objectives include formalizing stakeholder preferences, studying dynamic impacts, handling intersectionality, and designing transparent decision interfaces. Evaluation should combine mathematical analysis, user studies, historical data, and domain-specific governance criteria.
Applications and use cases