580.691 Learning Theory Reza Shadmehr Optimal feedback control stochastic feedback control with signal dependent noise.

Slides:



Advertisements
Similar presentations
Inventory Control.
Advertisements

Pattern Recognition and Machine Learning
Semidefinite Programming Machines
Positive and Negative Numbers
The Poisson distribution
Modeling of Data. Basic Bayes theorem Bayes theorem relates the conditional probabilities of two events A, and B: A might be a hypothesis and B might.
Tables, Figures, and Equations
Control and Feedback Introduction Open-loop and Closed-loop Systems
Chapter 6 The Mathematics of Diversification
Graphic Communication
Solving Systems of Linear Equations By Elimination
Bottoms Up Factoring. Start with the X-box 3-9 Product Sum
Array Operations ENGR 1181 MATLAB 4. Aerospace Engineers use turbulence data to calculate how close other planes can fly near the wake of a larger plane.
Powerpoint Jeopardy Category 1Category 2Category 3Category 4Category
16. Mean Square Estimation
Chapter 5 The Mathematics of Diversification
ECE 8443 – Pattern Recognition LECTURE 05: MAXIMUM LIKELIHOOD ESTIMATION Objectives: Discrete Features Maximum Likelihood Resources: D.H.S: Chapter 3 (Part.
1 Simple Linear Regression and Correlation The Model Estimating the Coefficients EXAMPLE 1: USED CAR SALES Assessing the model –T-tests –R-square.
Integration of sensory modalities
SYSTEMS Identification
MAE 552 Heuristic Optimization
Estimation and the Kalman Filter David Johnson. The Mean of a Discrete Distribution “I have more legs than average”
Kalman Filtering Pieter Abbeel UC Berkeley EECS Many slides adapted from Thrun, Burgard and Fox, Probabilistic Robotics TexPoint fonts used in EMF. Read.
Adaptive Signal Processing
Statistical learning and optimal control:
Learning Theory Reza Shadmehr Bayesian Learning 2: Gaussian distribution & linear regression Causal inference.
1 Pattern Recognition: Statistical and Neural Lonnie C. Ludeman Lecture 13 Oct 14, 2005 Nanjing University of Science & Technology.
ECE 8443 – Pattern Recognition LECTURE 06: MAXIMUM LIKELIHOOD AND BAYESIAN ESTIMATION Objectives: Bias in ML Estimates Bayesian Estimation Example Resources:
Population Coding Alexandre Pouget Okinawa Computational Neuroscience Course Okinawa, Japan November 2004.
Statistical learning and optimal control:
ECE 8443 – Pattern Recognition ECE 8423 – Adaptive Signal Processing Objectives: Deterministic vs. Random Maximum A Posteriori Maximum Likelihood Minimum.
Modern Navigation Thomas Herring
Homework: optimal control without a reference trajectory A disadvantage of the formulation that we developed in class is that it requires a reference trajectory.
ECE 8443 – Pattern Recognition LECTURE 10: HETEROSCEDASTIC LINEAR DISCRIMINANT ANALYSIS AND INDEPENDENT COMPONENT ANALYSIS Objectives: Generalization of.
Learning Theory Reza Shadmehr LMS with Newton-Raphson, weighted least squares, choice of loss function.
Statistical learning and optimal control: A framework for biological learning and motor control Lecture 4: Stochastic optimal control Reza Shadmehr Johns.
Learning Theory Reza Shadmehr Optimal feedback control stochastic feedback control with and without additive noise.
ECE 8443 – Pattern Recognition ECE 8423 – Adaptive Signal Processing Objectives: ML and Simple Regression Bias of the ML Estimate Variance of the ML Estimate.
CONSTANT EFFORT COMPUTATION AS A DETERMINANT OF MOTOR BEHAVIOR Emmanuel Guigon, Pierre Baraduc, Michel Desmurget INSERM U483, UPMC, Paris, France INSERM.
Solving Quadratic Equations Quadratic Equations: Think of other examples?
SYSTEMS Identification Ali Karimpour Assistant Professor Ferdowsi University of Mashhad Reference: “System Identification Theory For The User” Lennart.
Over-fitting and Regularization Chapter 4 textbook Lectures 11 and 12 on amlbook.com.
ECE 8443 – Pattern Recognition ECE 8527 – Introduction to Machine Learning and Pattern Recognition LECTURE 12: Advanced Discriminant Analysis Objectives:
Basic Business Statistics, 10e © 2006 Prentice-Hall, Inc. Chap 15-1 Chapter 15 Multiple Regression Model Building Basic Business Statistics 10 th Edition.
Joint Moments and Joint Characteristic Functions.
Section 1.7 Linear Independence and Nonsingular Matrices
Learning Theory Reza Shadmehr Distribution of the ML estimates of model parameters Signal dependent noise models.
Modern Control Systems (MCS) Dr. Imtiaz Hussain URL :
Solving Equations by Factoring.
The Quadratic Formula Quadratic Formula.
Learning Theory Reza Shadmehr Optimal control
Probability Theory and Parameter Estimation I
LECTURE 11: Advanced Discriminant Analysis
Solving Quadratic Equations by the Complete the Square Method
Factoring Quadratic Equations
Solving Equations by Factoring.
Zeros to Quadratic Functions
Integration of sensory modalities
Gaussian distribution & linear regression
Why is saccadic gain less than one?
Quadratic Equations.
Standard Form Quadratic Equation
Mathematical Foundations of BME
Learning Theory Reza Shadmehr
The loss function, the normal equation,
THE QUADRATIC FORMULA.
Mathematical Foundations of BME Reza Shadmehr
Mathematical Foundations of BME
Kalman Filter: Bayes Interpretation
Probabilistic Surrogate Models
Presentation transcript:

Learning Theory Reza Shadmehr Optimal feedback control stochastic feedback control with signal dependent noise

Representing signal dependent noise Vector of zero mean, variance 1 Gaussian random variables signal independent noisesignal dependent motor noise So the motor noise has mean zero and variance that grows with the square of the motor command.

Computing a cost for the motor commands: minimize endpoint variance Because there is noise in the motor commands, it will produce variance in our state. The above equation shows that the variance at the end of the movement is mostly influenced by the motor commands late in the movement. To see this, note that A is a matrix that when raised to a power, will become smaller. The larger the raised power, the smaller the resulting matrix will become. In the sum, we have a contribution from each motor command. When n is zero (the very first command), A is raised to a very high power. The noise in this command will have little influence on the endpoint variance. When n is larger (commands near end of the movement), A is raised to a small power. The noise in these commands will have a great deal of influence on the endpoint variance. Therefore, we have a natural cost function for the motor commands:

Cost per step: Control problem with signal dependent noise (Todorov, Neural Computation 2005)

Conjecture: If at some time point k+1 the value function under an optimal control policy is quadratic in x and e, and provided that we produce a u that minimizes the cost-to-go at time step k, then the value function at time step k will also be quadratic. To prove this, our first step is to find the u that minimizes the cost-to-go at time step k, and then show that at the resulting value function remains in the quadratic form above. To compute the expected value term, we need to do some work on the term e.

Terms that do not depend on u

So we just showed that if at some time point k+1 the value function under an optimal control policy is quadratic in x and e, and provided that we produce a u that minimizes the cost-to-go at time step k, then the value function at time step k will also be quadratic. Since we had earlier shown that at time step p-1 the cost is quadratic in x and e, we now have the solution to our problem.

Cost per step Summary: Control problem with signal dependent noise (Todorov 2005) For the last time step

Unlike the Gaussian noise, signal dependent noise affects the optimal control policy: feedback gain becomes smaller with increased signal dependent noise This reduction is particularly large near the end of the movement when the cost associated with motor commands tends to be larger sec s o P n i a G Variance of the motor noise Feedback gain for a 30 deg saccade