Rethinking Modelbased
Reinforcement Learning
Roberto Calandra
Facebook AI Research
UC Berkeley  23 October 2019
Reinforcement Learning as MDP
Reinforcement Learning Approaches
Modelfree:

Local convergence guaranteed*

Simple to implement

Computationally light

Does not generalize

Datainefficient
Modelbased:

No convergence guarantees

Challenging to learn model

Computationally intensive

Dataefficient

Generalize to new tasks
Evidence from neuroscience that humans use both approaches! [Daw et al. 2010]
Modelbased Reinforcement Learning
PILCO [Deisenroth et al. 2011]
PETS [Chua et al. 2018]
Chua, K.; Calandra, R.; McAllister, R. & Levine, S.
Deep Reinforcement Learning in a Handful of Trials using Probabilistic Dynamics Models
Advances in Neural Information Processing Systems (NIPS), 2018, 47544765
Experimental Results
Is Something Strange about MBRL?
How to Use the Reward?
GoalDriven Dynamics Learning
 Instead of optimizing the forward dynamics w.r.t. the NLL of the next state, we optimize w.r.t. the reward
(The reward is all we care about)
 Computing the gradients analytically is intractable
 We use a zeroorder optimizer: Bayesian optimization
 (and an LQG framework)
Bansal, S.; Calandra, R.; Xiao, T.; Levine, S. & Tomlin, C. J.
GoalDriven Dynamics Learning via Bayesian Optimization
IEEE Conference on Decision and Control (CDC), 2017, 51685173
Realworld Quadcopter
Dubins Car
Conclusion
There exist models that are wrong, but nearly optimal when used for control
 From a Sys.ID perspective, they are completely wrong
 These models might be outofclass (e.g., linear model for nonlinear dynamics)
 Hyphothesis: these models capture some structure of the optimal solution, ignoring the rest of the space
 Evidence: these models do not seem to generalize to new tasks
All models are wrong, but some are useful
 George E.P. Box
Very wrong models, can be very useful
 Roberto Calandra
If wrong models can be useful,
Can correct models be useless?
Model Likelihood vs Episode Reward
Objective Mismatch
Objective mismatch arises when one objective is optimized in the hope that a second, often uncorrelated, metric will also be optimized.
Negative LogLikelihood
Task Reward
Likelihood vs Reward
Deterministic model
Probabilistic model
Where is this assumption coming from?
Historical assumption ported from System Identification
Assumption: Optimizing the likelihood will optimize the reward
System Identification
Modelbased Reinforcement Learning
Sys ID vs MBRL
Objective Mismatch
Experimental results show that the likelihood of the trained models are not strongly correlated with task performance
What are the consequences?
Adversarially Generated Dynamics
What can we do?
 Modify objective when training dynamics model
 add controllability regularization [Singh et al. 2019]
 endtoend differentiable models [Amos et al. 2019]
 ...
 Move away from the singletask formulation
 ???
Reweighted Likelihood
How can we give more importance to data that are important for the specific task at hand?
Our attempt: reweight data w.r.t. distance from optimal trajectory
Overview
Introduced and analyzed Objective Mismatch in MBRL
 Identify fundamental flaw of current approach
 Provides new lens to understand MBRL
 Open exciting new venues
Lambert, N.; Amos, B.; Yadan, O. & Calandra, R.
Objective Mismatch in Modelbased Reinforcement Learning
Under review, Soon on Arxiv, 2019
If you are interested in collaborating, ping me
References
 Deisenroth, M.; Fox, D. & Rasmussen, C.
Gaussian Processes for DataEfficient Learning in Robotics and Control
IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), 2014, 37, 408423  Chua, K.; Calandra, R.; McAllister, R. & Levine, S.
Deep Reinforcement Learning in a Handful of Trials using Probabilistic Dynamics Models
Advances in Neural Information Processing Systems (NIPS), 2018, 47544765  Tian, S.; Ebert, F.; Jayaraman, D.; Mudigonda, M.; Finn, C.; Calandra, R. & Levine, S.
Manipulation by Feel: TouchBased Control with Deep Predictive Models
IEEE International Conference on Robotics and Automation (ICRA), 2019  Lambert, N.O.; Drew, D.S.; Yaconelli, J; Calandra, R.; Levine, S.; & Pister, K.S.J.
Low Level Control of a Quadrotor with Deep ModelBased Reinforcement Learning
IEEE Robotics and Automation Letters (RAL), 2019, 4, 42244230 
Bansal, S.; Calandra, R.; Xiao, T.; Levine, S. & Tomlin, C. J.
GoalDriven Dynamics Learning via Bayesian Optimization
IEEE Conference on Decision and Control (CDC), 2017, 51685173 
Lambert, N.; Amos, B.; Yadan, O. & Calandra, R.
Objective Mismatch in Modelbased Reinforcement Learning
Under review, Soon on Arxiv, 2019
Motivation
How to scale to more complex, unstructured domains?
Robotics
Learning to Fly a Quadcopter
Lambert, N.O.; Drew, D.S.; Yaconelli, J; Calandra, R.; Levine, S.; & Pister, K.S.J.
Low Level Control of a Quadrotor with Deep ModelBased Reinforcement Learning
IEEE Robotics and Automation Letters (RAL), 2019, 4, 42244230
Ablation Study
Design Choices in MBRL

Dynamics Model
 Forward dynamics
(Most used nowadays, since it is independent from the task and is causal, therefore allowing proper uncertainty propagation!)  What model to use? (Gaussian process, neural network, etc)
 Forward dynamics

How to compute longterm predictions?
 Usually, recursive propagation in the stateaction space
 Error compounds multiplicatively
 How do we propagate uncertainty?

What planner/policy to use?
 Training offline parametrized policy
 or using online Model Predictive Control (MPC)
Rethinking Modelbased Reinforcement Learning
By Roberto Calandra
Rethinking Modelbased Reinforcement Learning
[Berkeley]
 115