Research — Autonomous Systems·2018–2021
Decision-Making Under Uncertainty
Where the systems thinking started.
My role
NSF Graduate Research Fellow, Stanford
Outcome
Cut computational cost of control frameworks by ~50% while holding performance.
01
The problem
- Autonomous systems have to decide and act under uncertainty, in real time, with finite compute — often coordinating with other agents whose behavior they can't fully predict.
- The optimal decision is frequently the one you can't afford to compute in the time you have. The interesting problem is the trade-off, not the ideal.
02
The approach
- I designed end-to-end perception-to-control systems, balancing multiple predictive models against real-time constraints.
- I developed optimization and control frameworks — model predictive control, probabilistic cost functions — explicitly designed to be affordable, and used reinforcement learning and MDPs for adaptive, policy-based behavior.
- I built multi-agent coordination frameworks so distributed systems could make synchronized decisions.
03
The outcome
- Reduced the computational expense of the control frameworks by roughly half while maintaining performance.
- More durably: it shaped a way of thinking I still use — design for the best answer you can actually compute and trust, not the best answer in principle.
Technologies
Model Predictive ControlReinforcement LearningMDPsOptimizationMulti-Agent SystemsC++JuliaROS