Timezone: »

 
Spotlight
Lyapunov Density Models: Constraining Distribution Shift in Learning-Based Control
Katie Kang · Paula Gradu · Jason Choi · Michael Janner · Claire Tomlin · Sergey Levine

Tue Jul 19 11:35 AM -- 11:40 AM (PDT) @ Room 309
in Reinforcement Learning: Deep/Batch/Offline »

Learned models and policies can generalize effectively when evaluated within the distribution of the training data, but can produce unpredictable and erroneous outputs on out-of-distribution inputs. In order to avoid distribution shift when deploying learning-based control algorithms, we seek a mechanism to constrain the agent to states and actions that resemble those that the method was trained on. In control theory, Lyapunov stability and control-invariant sets allow us to make guarantees about controllers that stabilize the system around specific states, while in machine learning, density models allow us to estimate the training data distribution. Can we combine these two concepts, producing learning-based control algorithms that constrain the system to in-distribution states using only in-distribution actions? In this paper, we propose to do this by combining concepts from Lyapunov stability and density estimation, introducing Lyapunov density models: a generalization of control Lyapunov functions and density models that provides guarantees about an agent's ability to stay in-distribution over its entire trajectory.

Keywords: RL: Batch/Offline · RL: Risk Sensitive

Author Information

Katie Kang (UC Berkeley)
Paula Gradu (UC Berkeley)
Jason Choi (UC Berkeley)
Michael Janner (UC Berkeley)
Claire Tomlin (UC Berkeley)
Sergey Levine (University of California, Berkeley)

Related Events (a corresponding poster, oral, or spotlight)

More from the Same Authors