Current projects and collaborations

Stable gradient learning in continuous games

saddle Machine learning models with competing losses are notoriously challenging to train, e.g. zero-sum generative adversarial networks (GANs) or non-zero-sum reinforcement learning (RL) agents. We analyze and design learning algorithms used in these settings.

code, pa pe rs, sli des

Human/machine sensorimotor games

bpm In tasks involving adaptive algorithms and continuously-learning humans, rich dynamic behaviors can prevent the system from reaching consensus.

code, papers, sli des, video

Optimization interface for human autonomy

drone The aggressive dynamics of high-performance quadcopters demand agents to constantly be aware of the feedforward (planning) and feedback (reaction) policies of others.