Sam Burden currently: Postdoc in EECS at UC Berkeley Sep 2015: Asst Prof in EE at UW Seattle Dynamic Inverse Models in Human-Cyber-Physical Systems S.

Slides:



Advertisements
Similar presentations
Introductory Control Theory I400/B659: Intelligent robotics Kris Hauser.
Advertisements

David Rosen Goals  Overview of some of the big ideas in autonomous systems  Theme: Dynamical and stochastic systems lie at the intersection of mathematics.
1 Dynamic Programming Week #4. 2 Introduction Dynamic Programming (DP) –refers to a collection of algorithms –has a high computational complexity –assumes.
Adjoint Orbits, Principal Components, and Neural Nets Some facts about Lie groups and examples 2.Examples of adjoint orbits and a distance measure 3.Descent.
Properties of State Variables
Guaranteeing Safety in Semi-autonomous Robotic Systems: A Formal Approach through Hybrid Systems with Hidden Modes Domitilla Del Vecchio University of.
Robot Localization Using Bayesian Methods
Apprenticeship learning for robotic control Pieter Abbeel Stanford University Joint work with Andrew Y. Ng, Adam Coates, Morgan Quigley.
Quantifying Generalization from Trial-by-Trial Behavior in Reaching Movement Dan Liu Natural Computation Group Cognitive Science Department, UCSD March,
Model Predictive Control for Humanoid Balance and Locomotion Benjamin Stephens Robotics Institute.
INTRODUCTION TO Machine Learning ETHEM ALPAYDIN © The MIT Press, Lecture Slides for.
Organizational Notes no study guide no review session not sufficient to just read book and glance at lecture material midterm/final is considered hard.
1 AM3 Task 1.4 Stochastic Hybrid Models for Aerial and Ground Vehicles Sam Burden MAST Annual Review University of Pennsylvania March 8-9, 2010.
LEARNING FROM OBSERVATIONS Yılmaz KILIÇASLAN. Definition Learning takes place as the agent observes its interactions with the world and its own decision-making.
Pieter Abbeel and Andrew Y. Ng Apprenticeship Learning via Inverse Reinforcement Learning Pieter Abbeel Stanford University [Joint work with Andrew Ng.]
Pieter Abbeel and Andrew Y. Ng Apprenticeship Learning via Inverse Reinforcement Learning Pieter Abbeel Stanford University [Joint work with Andrew Ng.]
What is Cognitive Science? … is the interdisciplinary study of mind and intelligence, embracing philosophy, psychology, artificial intelligence, neuroscience,
Pieter Abbeel and Andrew Y. Ng Apprenticeship Learning via Inverse Reinforcement Learning Pieter Abbeel and Andrew Y. Ng Stanford University.
What is Cognitive Science? … is the interdisciplinary study of mind and intelligence, embracing philosophy, psychology, artificial intelligence, neuroscience,
Biological motor control Andrew Richardson McGovern Institute for Brain Research March 14, 2006.
Computational aspects of motor control and motor learning Michael I. Jordan* Mark J. Buller (mbuller) 21 February 2007 *In H. Heuer & S. Keele, (Eds.),
Learning and Planning for POMDPs Eyal Even-Dar, Tel-Aviv University Sham Kakade, University of Pennsylvania Yishay Mansour, Tel-Aviv University.
Definition of an Industrial Robot
The Reinforcement Learning Toolbox – Reinforcement Learning in Optimal Control Tasks Gerhard Neumann Master Thesis 2005 Institute für Grundlagen der Informationsverarbeitung.
CIS 540 Principles of Embedded Computation Spring Instructor: Rajeev Alur
Autumn 2008 EEE8013 Revision lecture 1 Ordinary Differential Equations.
Self-Organized Recurrent Neural Learning for Language Processing April 1, March 31, 2012 State from June 2009.
Machine Learning1 Machine Learning: Summary Greg Grudic CSCI-4830.
Modeling.
20/10/2009 IVR Herrmann IVR: Introduction to Control OVERVIEW Control systems Transformations Simple control algorithms.
NW Computational Intelligence Laboratory Implementing DHP in Software: Taking Control of the Pole-Cart System Lars Holmstrom.
10/6/2015 1Intelligent Systems and Soft Computing Lecture 0 What is Soft Computing.
UW Contributions: Past and Future Martin V. Butz Department of Cognitive Psychology University of Würzburg, Germany
Prediction in Human Presented by: Rezvan Kianifar January 2009.
Natural Actor-Critic Authors: Jan Peters and Stefan Schaal Neurocomputing, 2008 Cognitive robotics 2008/2009 Wouter Klijn.
Learning From Demonstration. Robot Learning A good control policy u=  (x,t) is often hard to engineer from first principles Reinforcement learning 
CS 478 – Tools for Machine Learning and Data Mining Backpropagation.
Dynamic Inverse Models in Cyber-Human Systems
1 Sequential Machine Theory Prof. K. J. Hintz Department of Electrical and Computer Engineering Lecture 1 Adaptation to this.
University of Windsor School of Computer Science Topics in Artificial Intelligence Fall 2008 Sept 11, 2008.
Feedback Linearization Presented by : Shubham Bhat (ECES-817)
Motor Control. Beyond babbling Three problems with motor babbling: –Random exploration is slow –Error-based learning algorithms are faster but error signals.
Structure and Synthesis of Robot Motion Topics in Control Subramanian Ramamoorthy School of Informatics 19 February, 2009.
Classification (slides adapted from Rob Schapire) Eran Segal Weizmann Institute.
Neural Networks Presented by M. Abbasi Course lecturer: Dr.Tohidkhah.
State Observer (Estimator)
Control and Synchronization of Chaos Li-Qun Chen Department of Mechanics, Shanghai University Shanghai Institute of Applied Mathematics and Mechanics Shanghai.
A paper1 by John Hauser, Shankar Sastry, and Petar Kokotovic
Chapter 8: Adaptive Networks

Adaptive Optimal Control of Nonlinear Parametric Strict Feedback Systems with application to Helicopter Attitude Control OBJECTIVES  Optimal adaptive.
Artificial Intelligence: Research and Collaborative Possibilities a presentation by: Dr. Ernest L. McDuffie, Assistant Professor Department of Computer.
Ferdinando A. Mussa-Ivaldi, “Modular features of motor control and learning,” Current opinion in Neurobiology, Vol. 9, 1999, pp The focus on complex.
Chapter 6 Neural Network.
A field of study that encompasses computational techniques for performing tasks that require intelligence when performed by humans. Simulation of human.
Chapter 15. Cognitive Adequacy in Brain- Like Intelligence in Brain-Like Intelligence, Sendhoff et al. Course: Robots Learning from Humans Cinarel, Ceyda.
CIS 540 Principles of Embedded Computation Spring Instructor: Rajeev Alur
Information Processing
Understanding Complex Systems May 15, 2007 Javier Alcazar, Ph.D.
CS 9633 Machine Learning Support Vector Machines
Cognitive Computing…. Computational Neuroscience
Predictive dynamical models for intermittent contact
An Overview of Reinforcement Learning
Introduction To Intelligent Control
Provably-safe interventions for Human-Cyber-Physical Systems (HCPS)
Intelligent Systems and
Chapter 8: Generalization and Function Approximation
Stability Analysis of Linear Systems
A Dynamic System Analysis of Simultaneous Recurrent Neural Network
Chapter 7 Inverse Dynamics Control
Presentation transcript:

Sam Burden currently: Postdoc in EECS at UC Berkeley Sep 2015: Asst Prof in EE at UW Seattle Dynamic Inverse Models in Human-Cyber-Physical Systems S. Shankar Sastry Dean of Engineering at UC Berkeley Embedded Humans: Provably Correct Decision Making for Networks of Humans and Unmanned Systems (N )

Increasingly mediated by automation – Augmented by hardware and software – Machines adapt to collaborate and assist Human interaction with the physical world Human-Cyber-Physical Systems (HCPS)

Embedding humans amid automation Can lead to performance degradation – pilot-induced oscillations in rotory/fixed-wing aircraft McRuer, Krendal 1974; Hess J. Guid. Cont. Dyn Pavel et al. Prog. Aero. Sci – overreliance on adaptive cruise control in cars Rudin-Brown, Parker Trans. Rch. F: Traffic Psych. and Behav Requires predictive models for human behavior

Predictable behavior from internal models Popular paradigm posits pairs of internal models – forward model predicts sensory effect of motor action Sutton, Barto Psych. Rev. 1981; Jordan, Rumehlart Cog. Sci. 1992; Wolpert, et al. Science 1995 – inverse model computes motor command expected to yield desired behavior Kawato Curr. Opin. Neurobio. 1999; Thoroughman, Shadmehr Nature 2000; Conditt, Mussa-Ivaldi PNAS 1999 – Theoretical and empirical evidence for paired forward + inverse models Bhushan, Shadmehr Bio. Cybern. 1999; Sanner, Kosha Bio. Cybern Hanuschkin, Ganguli, Hahnloser Front. Neural Circ. 2013; Giret, Kornfeld, Ganguli, Hahnloser PNAS 2014 Parallels in control theory, robotics, AI – Internal models, adaptive control, learning Francis, Wonham Automatica 1976; Sastry, Bodson Prentice Hall 1989; Sutton, Barto, Williams IEEE CSM 1992 Crawford, Sastry UCB EECS 1996; Atkeson, Schaal ICML 1997; Papavassiliou, Russell IJCAI 1999

Instantiating internal models Forward model ( M : U  Y ): static vs. dynamic – static map (linear or nonlinear) Hanuschkin, Ganguli, Hahnloser Front. Neural Circ Giret, Kornfeld, Ganguli, Hahnloser PNAS 2014 – dynamic map depends on intermediate state Thoroughman, Shadmehr Science 2000 Wolpert, Diedrichsen, Flanagan Nature Neurosci Inverse model ( M -1 : Y  U ): hard to define – static map may fail to be one-to-one or onto – dynamic map may be acausal or need state estimate

Today’s talk: dynamic inverse models from the perspective of mathematical control theory 1.derivation of dynamic inverse model 2.properties and implications for design of HCPS Dynamic inverse models in HCPS

Single input/single output forward model Consider forward model in control-affine form: – x in R n, u in R, y in R – f, g in C r (R n, R n ), h in C r (R n, R) Suppose model has strict relative degree  in N : – Expressed in terms of Lie derivatives L f h(x), L g h(x) : – intuitively, input affects  -th derivative of output – e.g.  =2 for Lagrangian mechanical systems applicable to interaction with physical world

Transformation of forward model Forward model: Suppose model has strict relative degree  in N : – e.g.  =2 for Lagrangian mechanical systems Then model is linear in new coordinates: – There exists such that in coordinates forward model has the form – Choosing yields simpler forward model

Dynamic inverse model Forward model: Given desired output y d, we seek desired input u d – Since y =  1, there exists unique v d such that – States  rendered unobservable by input v d ! Note that exact tracking is too stringent – need initial cond. But it’s easy to achieve exponential tracking – applying input yields  How does tracking affect unobservable states  ?

Tracking with stable model pair Forward model: Dynamic inverse model: Theorem: If forward and inverse models are exponentially stable, then feedforward input from dynamic inverse of internal model achieves exponential tracking for physical system. – Trajectories converge for stable model pairs (M, M -1 ) – Feedforward input “asymptotically inverts” dynamics

Tracking with stable model pair (M, M -1 ) Theorem implies: – For stable model pair, trajectories x, x’ converge to – Feedforward input “asymptotically inverts” dynamics (M, M -1 )

Suppose human ( H ) implements inverse model: – can infer desired task y H from observed input u H – nominal forward model becomes: – automation can intervene to improve performance by minimizing cost function J:R n  R using input  – guarantees performance improvement following intervention in human-cyber-physical system Application to provably-correct interventions

Today: predictive models for interaction Future: enhance human ability to interact with and control the built world – Human-Cyber-Physical – Human Intranet – Cybathlon Dynamics of humans embedded w/ machines Humans are the enabling technology

- Extensions and Generalizations - Properties of dynamic inverse model - Behavioral repertoire of humans Appendix

Extensions and generalizations Forward model: Dynamic inverse model: Results easily extend to accommodate: – multiple inputs / multiple outputs – (small) perturbations in dynamics Sastry Springer 1999 – approximate input-output linearization Hauser PhD Thesis 1989; Hauser, Sastry, Kokotovic IEEE TAC 1992; Banaszuk, Hauser SIAM JCO 1996 – learning / adaptation / estimation of dynamics Sutton, Barto, Williams IEEE CSM 1992; Papavassiliou, Russell ICJAI 1999 Sastry, Bodson Prentice Hall 1989; Vrabie, Vamvoudakis, Lewis IET 2013

Properties of dynamic inverse model Forward model: Dynamic inverse model: Property: dynamic inverse model is unique – Exact tracking input determined by y d – Independent of how internal model is represented or obtained (e.g. reinforcement learning, adaptive ctrl.) Sutton, Barto, Williams IEEE CSM 1992; Papavassiliou, Russell ICJAI 1999 Sastry, Bodson Prentice Hall 1989; Vrabie, Vamvoudakis, Lewis IET 2013 – Impossible to learn if inverse model is unstable

Behavioral repertoire of humans Too rich to model from first principles – Spans computational, algorithmic, & physical “levels of analysis” Marr, Poggio MIT AI MEMO 1976 – Influenced by neurophysiological state (cognitive load, hunger) LaPointe, Stierwalt, Maitland Int. J. Speech-Lang. Pathology 2010; Danziger, Levav, Avnaim-Pesso PNAS 2011 Can reduce dramatically during particular tasks – Bernstein posed the “problem of motor redundancy” Bernstein Pergamon Press – Perhaps instead we “exploit the bliss of motor abundance” e.g. using synergies, uncontrolled manifolds, optimality Latash Exp. Brain Rch. 2012; Ting, Macpherson J. Neurophys. 2005; Scholz, Schoner, Exp. Brain Rch Todorov, Jordan Nature Neurosci. 2002; Diedrichsen, Shadmehr, Ivry Trends Cog. Sci – For instance, locomotion naturally reduces dimensionality Burden, Revzen, Sastry IEEE TAC 2015