Performance measures
Regret, dynamic regret, competitive ratio, violation, refresh cost, information gain, and decision-aware freshness.
Research program
Mathematical foundations for reliable networked intelligence.
My research studies how intelligent systems can learn and coordinate under uncertainty while respecting safety, communication, computation, and reconfiguration realities.
reinforcement learningonline optimizationbanditsmulti-agent systemsnetwork optimization
Information
What can the learner know?
Partial state, delayed feedback, predictions, active probes, human preferences, shared experience, and their costs.
Dynamics
What changes over time?
Environments, workloads, threats, model versions, resource availability, and the value of information itself.
Coordination
What must work together?
Agents, actions, constraints, cloud–edge–device tiers, communication graphs, and coupled operational decisions.
Guarantees
What should remain true?
Safety at every step, low regret, bounded competitive loss, robust service, and scalable dependence on local structure.
01 · Safe & verifiable learning
Useful learning without a single unsafe action.
In safety-critical systems, average feasibility is not enough: a robot cannot collide once and compensate later, and jointly chosen actions may be unsafe even when every local action looks feasible.
Questions
- What can be learned from an initially certified safe seed?
- How do temporal viability and cross-component coupling change safe exploration?
- How can safety evidence be composed and returned before it becomes stale?
Ideas and techniques
- Conservative confidence certificates and robust predecessor operators.
- Optimism restricted to certified safe actions and protocols.
- Backward safe-set propagation, factor graphs, and composition-aware certification.
Target guaranteeszero instantaneous violationsublinear regretcertification complexitylocal-graph scalability
Representative work
02 · Prediction & costly adaptation
Adapt when change is valuable—not merely because change is possible.
Online decisions often have memory: moving a data-center workload, changing a traffic or NFV route, switching an antenna port, or refreshing a model consumes time, energy, bandwidth, and service continuity.
Questions
- How much can lookahead or prediction improve online performance?
- How should slow provisioning and fast scheduling be coupled?
- Can a method exploit good forecasts while remaining protected against bad ones?
Ideas and techniques
- Regularization with lookahead and competitive analysis.
- Switching, ramp, path-length, and reconfiguration cost models.
- Power-of-2, probe-then-commit, threshold control, and best-of-both-worlds learning.
Target guaranteescompetitive ratiodynamic regretforecast robustnessswitching-cost bounds
Representative work
03 · Partial & human information
Learn the state, the objective, and the value of asking.
In many systems neither the environment nor the objective is fully observed. State information may arrive selectively, feedback may be delayed, and human preferences may be multi-source, imperfect, or costly to query.
Questions
- When does partial online state revelation change POMDP learnability?
- What is the value of one extra probe, query, or shared trajectory?
- How can preference feedback support personalized and multi-objective decisions?
Ideas and techniques
- Belief/value decompositions and partial online state information.
- Active probing, proactive conversational queries, and probe-then-commit.
- Certified experience pooling, cross-action/state response sharing, information lower bounds, and RLHF models.
Target guaranteesregret & sample complexityinformation limitsbest-of-both-worlds sharingquery efficiency
Representative work
04 · Distributed & multi-agent intelligence
Use local structure to produce globally coherent intelligence.
Networked intelligence spans interacting agents and cloud, edge, and device tiers. The challenge is to coordinate learning and control without requiring full information, unlimited communication, or centralized execution.
Questions
- How can coupled agents coordinate safely using local observations and communication?
- When should experience, models, or model components be trained, tested, updated, or shared across tasks and tiers?
- How should communication, computation, memory, and control be co-designed?
Ideas and techniques
- Centralized training with communication-compatible decentralized execution.
- Local interaction graphs, message passing, and structured confidence sets.
- Certified pooling, task-aware memory, model partitioning, and Age-of-Intelligence signals.
Target guaranteeslocal-structure complexitycommunication efficiencycoordination safetytask-aware freshness
Representative work
05 · Dynamic, secure & robust AI
Preserve decision quality as the world—and the model—changes.
Nonstationary workloads, adversarial transitions, concept drift, stale models, LLM-security threats, and evolving attack surfaces make a one-time trained policy or static defense insufficient.
Questions
- Which changes are operationally consequential rather than merely statistically visible?
- When should an edge device refresh, a defender reallocate, or a learner switch regimes?
- What robustness is impossible without stronger feedback, structure, or assumptions?
Ideas and techniques
- Variation budgets, pathwise analysis, minimax RL, and regime-switching models.
- Age of Intelligence, KL and parameter certificates, and threshold synchronization.
- Queueing/risk dynamics, adaptive defense, and robustness lower bounds.
Target guaranteesminimax regretstale-decision / refresh tradeoffrobustness limitsrisk-aware defense
Representative work
Interactive research atlas
One project can contribute to several research directions.
Filter the portfolio to see how papers and current projects connect safety, information, adaptation, coordination, and robustness.
Mathematical toolkit
Different applications, recurring technical ideas.
Uncertainty control
Confidence sets, self-normalized concentration, martingales, robust optimization, and pessimistic certification.
Sequential structure
Bellman recursion, safe predecessor operators, post-decision states, variation budgets, and switching dynamics.
Information structure
POMDPs, active probing, preference queries, certified sharing, lower bounds, and measurement–disturbance tradeoffs.
Coordination structure
Factor graphs, local scopes, centralized training/decentralized execution, message passing, and cross-level optimization.
Impossibility and limits
Information-theoretic lower bounds, safety margins, exponential joint spaces, robustness barriers, and non-identifiability.
Research collaboration
Interested in a shared problem across theory and intelligent systems?
Potential collaborations include safe and multi-agent learning, online optimization, edge and wireless intelligence, AI security, foundation models, robotics, and quantum networking.