Near-Optimal Decision-Making in Dynamic Environments Manu Chhabra 1 Robert Jacobs 2 1 Department of Computer Science 2 Department of Brain & Cognitive.

Slides:



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

ABS Control Project Ondrej Ille Pre-bachelor Project.
Yasuhiro Fujiwara (NTT Cyber Space Labs)
Mechatronics 1 Weeks 5,6, & 7. Learning Outcomes By the end of week 5-7 session, students will understand the dynamics of industrial robots.
Kinetic Rules Underlying Multi-Joint Reaching Movements. Daniel M Corcos†, James S. Thomas*, and Ziaul Hasan†. School of Physical Therapy*, Ohio University,
Resource Management of Highly Configurable Tasks April 26, 2004 Jeffery P. HansenSourav Ghosh Raj RajkumarJohn P. Lehoczky Carnegie Mellon University.
CSCE 641: Forward kinematics and inverse kinematics Jinxiang Chai.
Quantifying Generalization from Trial-by-Trial Behavior in Reaching Movement Dan Liu Natural Computation Group Cognitive Science Department, UCSD March,
ME 4135 Fall 2011 R. R. Lindeke, Ph. D. Robot Dynamics – The Action of a Manipulator When Forced.
Optimizing number of hidden neurons in neural networks
INVESTMENTS | BODIE, KANE, MARCUS ©2011 The McGraw-Hill Companies CHAPTER 7 Optimal Risky Portfolios 1.
CSCE 641: Forward kinematics and inverse kinematics Jinxiang Chai.
Integrating POMDP and RL for a Two Layer Simulated Robot Architecture Presented by Alp Sardağ.
Chapter 3: The Efficiency of Algorithms Invitation to Computer Science, C++ Version, Fourth Edition.
CSCE 689: Forward Kinematics and Inverse Kinematics
Single Point of Contact Manipulation of Unknown Objects Stuart Anderson Advisor: Reid Simmons School of Computer Science Carnegie Mellon University.
Mechatronics 1 Week 9 & 10. Learning Outcomes By the end of week 9-10 session, students will understand the control system of industrial robots.
Trends in Motor Control
Dimitrios Konstantas, Evangelos Grigoroudis, Vassilis S. Kouikoglou and Stratos Ioannidis Department of Production Engineering and Management Technical.
Definition of an Industrial Robot
Asaf Cohen (joint work with Rami Atar) Department of Mathematics University of Michigan Financial Mathematics Seminar University of Michigan March 11,
Contrasts and Basis Functions Hugo Spiers Adam Liston.
Learning: Nearest Neighbor Artificial Intelligence CMSC January 31, 2002.
T for Two: Linear Synergy Advances the Evolution of Directional Pointing Behaviour Marieke Rohde & Ezequiel Di Paolo Centre for Computational Neuroscience.
Chen Cai, Benjamin Heydecker Presentation for the 4th CREST Open Workshop Operation Research for Software Engineering Methods, London, 2010 Approximate.
20/10/2009 IVR Herrmann IVR: Introduction to Control OVERVIEW Control systems Transformations Simple control algorithms.
Robot Dynamics – Slide Set 10 ME 4135 R. R. Lindeke, Ph. D.
Dynamics.  relationship between the joint actuator torques and the motion of the structure  Derivation of dynamic model of a manipulator  Simulation.
STUDY, MODEL & INTERFACE WITH MOTOR CORTEX Presented by - Waseem Khatri.
Chapter 5 Trajectory Planning 5.1 INTRODUCTION In this chapters …….  Path and trajectory planning means the way that a robot is moved from one location.
Chapter 5 Trajectory Planning 5.1 INTRODUCTION In this chapters …….  Path and trajectory planning means the way that a robot is moved from one location.
Natural Actor-Critic Authors: Jan Peters and Stefan Schaal Neurocomputing, 2008 Cognitive robotics 2008/2009 Wouter Klijn.
ANTs PI Meeting, Nov. 29, 2000W. Zhang, Washington University1 Flexible Methods for Multi-agent distributed resource Allocation by Exploiting Phase Transitions.
T. Bajd, M. Mihelj, J. Lenarčič, A. Stanovnik, M. Munih, Robotics, Springer, 2010 ROBOT CONTROL T. Bajd and M. Mihelj.
Angular Kinetics After reading this chapter, the student should be able to: Define torque and discuss the characteristics of a torque. State the angular.
12 November 2009, UT Austin, CS Department Control of Humanoid Robots Luis Sentis, Ph.D. Personal robotics Guidance of gait.
Driveline Dynamics Engine Dynamics Driveline and Efficiency Gearbox and Clutch Dynamics Gearbox Design.
A Power Grid Analysis and Verification Tool Based on a Statistical Prediction Engine M.K. Tsiampas, D. Bountas, P. Merakos, N.E. Evmorfopoulos, S. Bantas.
Control 1 Keypoints: The control problem Forward models: –Geometric –Kinetic –Dynamic Process characteristics for a simple linear dynamic system.
Basic Biomechanical Factors and Concepts
Progress in identification of damping: Energy-based method with incomplete and noisy data Marco Prandina University of Liverpool.
CSCE 441: Computer Graphics Forward/Inverse kinematics Jinxiang Chai.
Motor Control. Beyond babbling Three problems with motor babbling: –Random exploration is slow –Error-based learning algorithms are faster but error signals.
Value Function Approximation on Non-linear Manifolds for Robot Motor Control Masashi Sugiyama1)2) Hirotaka Hachiya1)2) Christopher Towell2) Sethu.
Cognition – 2/e Dr. Daniel B. Willingham
The Restricted Matched Filter for Distributed Detection Charles Sestok and Alan Oppenheim MIT DARPA SensIT PI Meeting Jan. 16, 2002.
Accurate Robot Positioning using Corrective Learning Ram Subramanian ECE 539 Course Project Fall 2003.
HUMAN FACTORS FOR INTERIOR ARCHITECTURE
Limits On Wireless Communication In Fading Environment Using Multiple Antennas Presented By Fabian Rozario ECE Department Paper By G.J. Foschini and M.J.
1 Optimization Techniques Constrained Optimization by Linear Programming updated NTU SY-521-N SMU EMIS 5300/7300 Systems Analysis Methods Dr.
MURI High- Level Control Biomimetic Robots - ONR Site Visit - August 9, 2000 Human Computational Modeling PurposePurpose: to understand arm impedance.
Λ-Model and Equilibrium Point Hypothesis References: 1.Latash M.L., Control of human movement, chapters 1-3, Human kinetics Publishers, Feldman.
MURI High-Level Control Biomimetic Robots - ONR Site Visit - August 9, Fabrication MURI Low-Level Control High-Level Control What strategies are.
Discovering Optimal Training Policies: A New Experimental Paradigm Robert V. Lindsey, Michael C. Mozer Institute of Cognitive Science Department of Computer.
CSCE 441: Computer Graphics Forward/Inverse kinematics Jinxiang Chai.
SPARSITY & SPEECH SCIENCE? TOWARDS A DATA-DRIVEN CHARACTERIZATION OF SPEECH MOTOR CONTROL V IKRAM R AMANARAYANAN University of Southern California, Los.
Rotational Dynamics Rode, Kiana, Tiana, and Celina.
Virtual Gravity Control for Swing-Up pendulum K.Furuta *, S.Suzuki ** and K.Azuma * * Department of Computers and Systems Engineering, TDU, Saitama Japan.
Real-time Simulation and Processing of Peripheral Nerve Spike Activity David Wallace Croft Presented to the Dallas Area Neuroscience.
CSCE 441: Computer Graphics Forward/Inverse kinematics
Introducing the M-metric Maurice R. Masliah and Paul Milgram
Analytics and OR DP- summary.
Accurate Robot Positioning using Corrective Learning
Dynamical Models of Decision Making Optimality, human performance, and principles of neural information processing Jay McClelland Department of Psychology.
Forward & Backward selection in hybrid network
Chapter 3: The Efficiency of Algorithms
CSCE 441: Computer Graphics Forward/Inverse kinematics
Chapter 3: The Efficiency of Algorithms
Chapter 4 . Trajectory planning and Inverse kinematics
Presentation transcript:

Near-Optimal Decision-Making in Dynamic Environments Manu Chhabra 1 Robert Jacobs 2 1 Department of Computer Science 2 Department of Brain & Cognitive Sciences University of Rochester

Dynamic Decision-Making Decision-making in environments with complex temporal dynamics –Decision-making at many moments in time –Temporal dependencies among decisions Examples: –Flying an airplane –Piloting a boat –Controlling an industrial process –Coordinating firefighters to fight a fire

Outline Experimental project: –Is human adaptive control optimal across different noise environments? Computational project: –Can optimal movements be planned as linear combinations of optimal motor primitives?

Dynamics and Noise Adaptive control requires learning about both the dynamics and the noise of a complex system Dynamics: relationship between control signals and the expected responses to these signals Noise: relationship between control signals and the variances of the responses to these signals

Dynamics and Noise Dynamics: 2 nd –order linear system –Object position, velocity, acceleration: –Mass: m –Force: f –Viscous resistance: b Noise: corrupts force f

Three Noise Conditions No-Noise (NN) Proportional Noise (PN) –Small forces are corrupted by small amounts of noise –Large forces are corrupted by large amounts of noise Inversely-Proportional Noise (IPN) –Small forces are corrupted by large amounts of noise –Large forces are corrupted by small amounts of noise

Ideal Actors Optimal control laws computed via dynamic programming –Optimal control law depends on the noise characteristics of the environment –Different ideal actors were created for different noise conditions Efficiency: –Ratio of subject’s performance to expected performance of ideal actor

Experimental Results

Proportional NoiseInversely-Proportional Noise Ideal Actor Average over subjects

Conclusions Subjects learned control strategies tailored to the specific noise characteristics of their conditions –Allowed them to achieve levels of performance near the information-theoretic upper bounds Conclude: Subjects learned to efficiently use all available information to plan and execute control policies that maximized performances on their tasks

Conclusions Q: Is human adaptive control optimal across different noise environments? A: Yes (under the conditions studied here)

Computational Complexity of Motor Control Task: Apply torques to a two-joint arm so that its endpoint moves from location A to location B in 100 time steps Assume: At each moment in time, torque is either on or off at each joint Q: How many torque sequences are possible solutions? A: “Curse of dimensionality”

Motor Synergies Motor synergies: dependencies among degrees of freedom Motor synergies = motor primitives –Basic units of behavior that can be linearly combined to form complex units of behavior –To form complex behavior: only need to specify linear coefficients Behavioral and physiological evidence

Approach Hypothesis: Optimal motor control can be achieved by combining a small number of scaled and time-shifted optimal synergies If so, motor control is easy –Only need to specify scaling coefficients and time-shifts Q: How do we find optimal synergies?

Strategy First, find optimal solutions to tasks in training set –Optimal solution is an optimal sequence of torques that moves a motor system from an initial state to a goal state Next, perform dimensionality reduction on space of possible solutions –Optimal solutions lie on a low-dimensional manifold –Important directions = motor synergies –Technique: non-negative matrix factorization

Strategy Lastly, find solutions to novel tasks in test set using synergies –Linear coefficients –Time-shifts

Motor Tasks Reaching task: move the endpoint of a simulated two-joint robot arm from one location to another in a specified time period Via-point task: move from one location to another while passing through an intermediate location

Simulations Example: Reaching task 256 tasks in training set –Find (approximate) optimal solutions to each task –Find optimal motor synergies via dimensionality reduction 64 tasks in test set –Find solution to each task by combining motor synergies Linear coefficients Time-shifts

How Many Synergies Are Needed? Reaching taskVia-Point task

Task-Dependent vs. Task-Independent Synergies

Synergies from Reaching Task

Synergies from Via-Point Task

Fast Learning with Synergies

Summary Optimal solutions lie on a low-dimensional manifold – Dimensionality reduction for discovering optimal synergies Near-optimal motor control by combining scaled and time- shifted synergies A small number of synergies are sufficient Task-dependent and task-independent synergies Learning with synergies is fast Additional research: two-joint arm with muscle model

Future Directions ??? Normative Ideal Actor : –unlimited computational power –unlimited memory –Provides information-theoretic upper bound on performance Human Ideal Actor: –limited computational power –limited working and long-term memory –Provides upper bound on performance if one has human cognitive limitations

Experimental Results

Dimensionality Reduction