Neuro-Adaptive Design - I:

Similar documents
Neuro-Adaptive Design II:

Chapter 5. Solution of System of Linear Equations. Module No. 6. Solution of Inconsistent and Ill Conditioned Systems

Errors for Linear Systems

CHAPTER 5 NUMERICAL EVALUATION OF DYNAMIC RESPONSE

Zeros and Zero Dynamics for Linear, Time-delay System

LINEAR REGRESSION ANALYSIS. MODULE IX Lecture Multicollinearity

Introduction. - The Second Lyapunov Method. - The First Lyapunov Method

Module 3 LOSSY IMAGE COMPRESSION SYSTEMS. Version 2 ECE IIT, Kharagpur

Econ107 Applied Econometrics Topic 3: Classical Model (Studenmund, Chapter 4)

8.4 COMPLEX VECTOR SPACES AND INNER PRODUCTS

CHAPTER III Neural Networks as Associative Memory

Some Comments on Accelerating Convergence of Iterative Sequences Using Direct Inversion of the Iterative Subspace (DIIS)

Outline. Communication. Bellman Ford Algorithm. Bellman Ford Example. Bellman Ford Shortest Path [1]

APPENDIX A Some Linear Algebra

Marginal Effects in Probit Models: Interpretation and Testing. 1. Interpreting Probit Coefficients

Lectures - Week 4 Matrix norms, Conditioning, Vector Spaces, Linear Independence, Spanning sets and Basis, Null space and Range of a Matrix

Robust observed-state feedback design. for discrete-time systems rational in the uncertainties

4DVAR, according to the name, is a four-dimensional variational method.

Week 5: Neural Networks

State Estimation. Ali Abur Northeastern University, USA. September 28, 2016 Fall 2016 CURENT Course Lecture Notes

Linear Feature Engineering 11

Hidden Markov Models & The Multivariate Gaussian (10/26/04)

Dr. Shalabh Department of Mathematics and Statistics Indian Institute of Technology Kanpur

Design and Optimization of Fuzzy Controller for Inverse Pendulum System Using Genetic Algorithm

Hongyi Miao, College of Science, Nanjing Forestry University, Nanjing ,China. (Received 20 June 2013, accepted 11 March 2014) I)ϕ (k)

Polynomial Regression Models

Feature Selection: Part 1

Singular Value Decomposition: Theory and Applications

Quantum Mechanics I - Session 4

Problem Set 9 Solutions

Inner Product. Euclidean Space. Orthonormal Basis. Orthogonal

Lecture Notes on Linear Regression

Report on Image warping

Support Vector Machines. Vibhav Gogate The University of Texas at dallas

EEE 241: Linear Systems

Generalized Linear Methods

High resolution entropy stable scheme for shallow water equations

Computing Correlated Equilibria in Multi-Player Games

Solution of Linear System of Equations and Matrix Inversion Gauss Seidel Iteration Method

PARTICIPATION FACTOR IN MODAL ANALYSIS OF POWER SYSTEMS STABILITY

VQ widely used in coding speech, image, and video

COMPARISON OF SOME RELIABILITY CHARACTERISTICS BETWEEN REDUNDANT SYSTEMS REQUIRING SUPPORTING UNITS FOR THEIR OPERATIONS

Parametric fractional imputation for missing data analysis. Jae Kwang Kim Survey Working Group Seminar March 29, 2010

A Computational Viewpoint on Classical Density Functional Theory

LINEAR REGRESSION ANALYSIS. MODULE VIII Lecture Indicator Variables

Digital Signal Processing

2 STATISTICALLY OPTIMAL TRAINING DATA 2.1 A CRITERION OF OPTIMALITY We revew the crteron of statstcally optmal tranng data (Fukumzu et al., 1994). We

Notes on Frequency Estimation in Data Streams

On Event-Triggered Adaptive Architectures for Decentralized and Distributed Control of Large- Scale Modular Systems

LINEAR REGRESSION ANALYSIS. MODULE IX Lecture Multicollinearity

On a direct solver for linear least squares problems

Convexity preserving interpolation by splines of arbitrary degree

Dynamic Systems on Graphs

Maximum Likelihood Estimation of Binary Dependent Variables Models: Probit and Logit. 1. General Formulation of Binary Dependent Variables Models

ADVANCED MACHINE LEARNING ADVANCED MACHINE LEARNING

hapter 6 System Norms 6. Introducton s n the matrx case, the measure on a system should be nduced from the space of sgnals t operates on. hus, the sze

Operating conditions of a mine fan under conditions of variable resistance

Module 2. Random Processes. Version 2 ECE IIT, Kharagpur

Robust Sliding Mode Observers for Large Scale Systems with Applications to a Multimachine Power System

Lecture 21: Numerical methods for pricing American type derivatives

LOW BIAS INTEGRATED PATH ESTIMATORS. James M. Calvin

A RELAXED SUFFICIENT CONDITION FOR ROBUST STABILITY OF AFFINE SYSTEMS

Numerical Heat and Mass Transfer

CinChE Problem-Solving Strategy Chapter 4 Development of a Mathematical Model. formulation. procedure

2. PROBLEM STATEMENT AND SOLUTION STRATEGIES. L q. Suppose that we have a structure with known geometry (b, h, and L) and material properties (EA).

Application of B-Spline to Numerical Solution of a System of Singularly Perturbed Problems

Markov Chain Monte Carlo (MCMC), Gibbs Sampling, Metropolis Algorithms, and Simulated Annealing Bioinformatics Course Supplement

Note 10. Modeling and Simulation of Dynamic Systems

Off-policy Reinforcement Learning for Robust Control of Discrete-time Uncertain Linear Systems

Adaptive sliding mode reliable excitation control design for power systems

Analytical Gradient Evaluation of Cost Functions in. General Field Solvers: A Novel Approach for. Optimization of Microwave Structures

Additional Codes using Finite Difference Method. 1 HJB Equation for Consumption-Saving Problem Without Uncertainty

Linear, affine, and convex sets and hulls In the sequel, unless otherwise specified, X will denote a real vector space.

Outline and Reading. Dynamic Programming. Dynamic Programming revealed. Computing Fibonacci. The General Dynamic Programming Technique

Calculation of time complexity (3%)

A New Refinement of Jacobi Method for Solution of Linear System Equations AX=b

FTCS Solution to the Heat Equation

PREDICTIVE CONTROL BY DISTRIBUTED PARAMETER SYSTEMS BLOCKSET FOR MATLAB & SIMULINK

Erratum: A Generalized Path Integral Control Approach to Reinforcement Learning

The Ordinary Least Squares (OLS) Estimator

Matrix Approximation via Sampling, Subspace Embedding. 1 Solving Linear Systems Using SVD

College of Computer & Information Science Fall 2009 Northeastern University 20 October 2009

THE SUMMATION NOTATION Ʃ

Lecture 4: Constant Time SVD Approximation

Chapter 9: Statistical Inference and the Relationship between Two Variables

Observers for non-linear differential-algebraic. system

Probability Theory (revisited)

Number of cases Number of factors Number of covariates Number of levels of factor i. Value of the dependent variable for case k

Quantifying Uncertainty

10.34 Numerical Methods Applied to Chemical Engineering Fall Homework #3: Systems of Nonlinear Equations and Optimization

CME 302: NUMERICAL LINEAR ALGEBRA FALL 2005/06 LECTURE 13

Perron Vectors of an Irreducible Nonnegative Interval Matrix

Lecture 12: Classification

Changing Topology and Communication Delays

Lecture 2 Solution of Nonlinear Equations ( Root Finding Problems )

A PROCEDURE FOR SIMULATING THE NONLINEAR CONDUCTION HEAT TRANSFER IN A BODY WITH TEMPERATURE DEPENDENT THERMAL CONDUCTIVITY.

Linear Approximation with Regularization and Moving Least Squares

Improved delay-dependent stability criteria for discrete-time stochastic neural networks with time-varying delays

International Journal of Pure and Applied Sciences and Technology

Week 2. This week, we covered operations on sets and cardinality.

Transcription:

Lecture 36 Neuro-Adaptve Desgn - I: A Robustfyng ool for Dynamc Inverson Desgn Dr. Radhakant Padh Asst. Professor Dept. of Aerospace Engneerng Indan Insttute of Scence - Bangalore

Motvaton Perfect system modelng s dffcult Sources of mperfecton: Unmodelled dynamcs (mssng algebrac terms n the model) Inaccurate knowledge of system parameters and/or change of system parameters durng operaton Inaccuracy n computatons (lke matrx nverson) Objectve: o ncrease the robustness of dynamc nverson wth respect to parameter and/or modelng naccuraces Dr. Radhakant Padh, AE Dept., IISc-Bangalore

Reference B. S. Km and A. J. Calse, Nonlnear Flght Control Usng Neural Networks, Journal of Gudance, Control and Dynamcs, Vol. 0, No. 1, 1997, pp. 6-33. Dr. Radhakant Padh, AE Dept., IISc-Bangalore 3

Neuro-adaptve desgn System Assumptons: Goal: () ( 3 ) X = f X X (,, δ ) n X, X R, δ R 1 X, X m are avalable for control computaton m= n X, X, δ R f and X (square system,.e. ) s nvertble X X X X c c ( : commanded sgnal ) c m Dr. Radhakant Padh, AE Dept., IISc-Bangalore 4

Neuro-adaptve desgn Ideal Case: Defne hen δ X X X c Desgn such that X + K X + K X = 0, where K, K > 0 pdf matrces d p p d, If K = dag k, K = dag k, k k > 0 d d p p p d x + k x + k x = d p c d p 0 x x + k x + k x = 0 Dr. Radhakant Padh, AE Dept., IISc-Bangalore 5

Neuro-adaptve desgn x = x + k x + k x c d p u th ( compoent of "pseudo control") Repeatng ths exercse for = 1,,..., n we get X = U ( U : Pseudo control varable) f X X = U (,, δ ) δ = f 1 ( X, X, U) wth the assumton that the nverse exts Dr. Radhakant Padh, AE Dept., IISc-Bangalore 6

Neuro-adaptve desgn Note: Real - tme computaton of ths nverse may be dffcult. In the case, a Neural Network can learn ths relatonshp offlne (n an approxmate sense) 1 = f ( X X U ) NN 1 X, X, U.e, ˆ δ,, = Problem: he model and the neural network tranng are NO perfect...these ntroduce errors. Dr. Radhakant Padh, AE Dept., IISc-Bangalore 7

Neuro-adaptve desgn Soluton : Desgn another neural network to cancel ths error! Desgn of Adaptve NN After syntheszng, the system dynamcs s (,, ˆ δ ) X = f X X ˆ δ = f ( X, X, δ ) + f ( X, X, ˆ δ) f ( X, X, δ) U U Δ ( X, X, U) Dr. Radhakant Padh, AE Dept., IISc-Bangalore 8

Neuro-adaptve desgn.e X = U +Δ X, X, U Dr. Radhakant Padh, AE Dept., IISc-Bangalore 9

Neuro-adaptve desgn Note: If Δ=, then the error dynamcs behaves lke the deal case, and hence, the objectve wll be met. rck : Modfy the Pseudo control as where 0 ˆ and : U ad U U = U ˆ I U U = X + K X + K X I c d p s the P seudo control for the "deal case". Adaptve control to cancel the unwarted effect. ad Note: Pseudo-control modfcaton s the reason why technque s applcable to dynamc nverson only. Dr. Radhakant Padh, AE Dept., IISc-Bangalore 10

Neuro-adaptve desgn Wth ths, the closed loop error dynamcs becomes X + K X + K X = X X + K X + K X.e, d p c d p ( X ) (,, ) c KdX KpX U X X U = + + +Δ = U Queston Can we desgn I ( U ) I Uad Δ X X U ( X X U) ˆ,, X + K ˆ dx + KpX = Uad Δ X, X, U : Uˆ Uˆ Δ,, ad ad (adaptvely ) such that Dr. Radhakant Padh, AE Dept., IISc-Bangalore 11

Neuro-adaptve desgn If ths happens, then th Defne channel : X, X 0 (approxmately) x + k x + k x = Uˆ Δ X, X, U e d p ad x x 0 1 0 hen e ˆ = e + Uad, X kp k Δ 1 d A b XU, Dr. Radhakant Padh, AE Dept., IISc-Bangalore 1

Neuro-adaptve desgn Goal for Uˆ ad Uˆ he "deal purpose" of s to capture the functon Δ "pefectly" so that e 0 asymptotcally However, ths s dffcult to acheve. Hence, am for "practcal stablty "only.e, e ad remans bounded, where the bound can be made arbtrarty small. Dr. Radhakant Padh, AE Dept., IISc-Bangalore 13

Neuro-adaptve desgn Let Δˆ,, be a NN realzaton of,, ( X XU) Δ X XU when a fnte number ( N) of bass functons β X, X, U, j = 1,,..., N are used, wth the j correspondng weghts of the network beng () Wˆ t, j = 1,,..., N j.e, ˆ ˆ ( X, X, U) W () t β ( X, X, U) Δ = j j Dr. Radhakant Padh, AE Dept., IISc-Bangalore 14

Neuro-adaptve desgn β 1 β ˆ Δ (,, ) ˆ, ˆ,..., ˆ X X U = W W 1 W N... β N ( ) () β ( ).e. Uˆ = Δ ˆ X, X, U = Wˆ t. X, X, U Ideal case: ad { X X U} ( ) () β ( ) When Wˆ s optmzed over some compact doman n,,, * * * let the result be Wˆ. In that case Δ ˆ X, X, U = Wˆ t. X, X, U Δ Δ ˆ ˆ < ε ε * and where s the deal error of functon. Dr. Radhakant Padh, AE Dept., IISc-Bangalore 15

Neuro-adaptve desgn Note: ε depends on the followngs: () N : Sze of the network () Choce of bass functons β ( j = 1,..., N) () Accuracy of the frst neural network th (whch determnes Δ ) j Dr. Radhakant Padh, AE Dept., IISc-Bangalore 16

Neuro-adaptve desgn Estmaton Error: () = ˆ () * () * β,, β (,, ) ˆ ˆ ˆ ˆ U Δ = W t X XU W X XU ad * () ˆ β ( ) = Wˆ t W X, X, U = W β,, ( X X U) W t W t Wˆ where, : Error n weght at tme Note: Wˆ s constant. Hence = = W * () t Wˆ () t Wˆ ˆ W() t = 0 * t Dr. Radhakant Padh, AE Dept., IISc-Bangalore 17

Neuro-adaptve desgn Notaton : β ( X, X, U) = β (,, ) t e W Δ ˆ =Δˆ ( X, X, U) (,, ) t e W * * Δ ˆ ˆ =Δ ( X, X, U) t, e, W Used n mplementaton X x Note: e, e X x used n proof Dr. Radhakant Padh, AE Dept., IISc-Bangalore 18

Neuro-adaptve desgn th Error dynamcs n channel : Note : ( ˆ ) e = Ae + b U Δ ad ( ˆ ˆ * * ˆ ) = Ae + b U Δ +Δ Δ ad = Ae + bw β t e W + b Δ Δ ( *,, ˆ ) A 0 1 = s a Hurwtz matrx. kp k d Characterstc equaton: s + k s+ k = 0, wth k, k > 0 d p d p Dr. Radhakant Padh, AE Dept., IISc-Bangalore 19

Neuro-adaptve desgn Stablty Analyss: Defne a Lyapunov Functon canddate: ( ) where, V e, W V = n = 1 (, ) V e W 1 1 e Pe + W W, e > E P γ = 1 E + W W, e E E : defnes "dead zone" P γ By defnton, e e Pe P s a pdf matrx satsfyng P PA + A P = Q, Q > 0 (pdf) Note : By the selecton, V s contnuous at the boundary of dead zone. Dr. Radhakant Padh, AE Dept., IISc-Bangalore 0

Neuro-adaptve desgn Note: (1) Exstence of such a pdf matrx P s guaranteed by the fact A s Hurwtz. P () If Q = I, then the soluton for s gven by kd k 1 p 1 + 1 + kp k d k p k p = 1 1 1 1+ kp k d k p Dr. Radhakant Padh, AE Dept., IISc-Bangalore 1

Neuro-adaptve desgn 1 1 V = ( e Pe + e Pe ) + W W γ ˆ * ( ˆ * β,, ) Ae + bw β t, e, W b Pe + Δ Δ 1 = + WW ˆ e P Ae bw t e W b γ + + + Δ Δ 1 ( ˆ * ) 1 = e A P + PA e + e Pb W β + Δ Δ + W Wˆ γ 1 ˆ W e Qe + e Pb ε + W e Pb β + γ Weght update rule: Wˆ = γ e Pbβ Dr. Radhakant Padh, AE Dept., IISc-Bangalore

Neuro-adaptve desgn 1 e Qe + ε e P Pb P P 1 e λmn ( Q ) + ε e P P b λmax However, e Pe P e e Pe.e, e λ max ( P ) e Pe e = λ max ( s dened snce s pdf ) ( P) λ ( P) e max P Dr. Radhakant Padh, AE Dept., IISc-Bangalore 3

Neuro-adaptve desgn 1 e V Q e P Pe P P λmn + ε λmax λmax ( P ) 1 e λmn ( Q ) P = + ε e Pe λmax ( P ) e P ( Q) ( P ) ( Q) ( P ) ( P) max 1 e λ V e P P mn + ε λ P max λmax 1 e λ V 0, when + < 0 ( P) P mn ε λmax λmax λ Dr. Radhakant Padh, AE Dept., IISc-Bangalore 4

Neuro-adaptve desgn.e, ε λ max P < 1 e P λ λ max mn P Q ().e, e P > ε λ λ mn max Note: 1 hs defnes E, However t contans ε, Q whch s unkown. However, t may requre teratve smulaton study. P 3 Dr. Radhakant Padh, AE Dept., IISc-Bangalore 5

Neuro-adaptve desgn If Q = I ( P) E= ε λmax where, P s gven before. ( 3 ) Selectng Q = I also leads to the least bounds. Insde the dead zone: V = W Wˆ 3 Select Wˆ = 0 hen V = 0 Note: By the selecton, V s contnuous at the boundary of ths deadzone. Dr. Radhakant Padh, AE Dept., IISc-Bangalore 6

Neuro-adaptve Desgn: Implementaton of Controller (1) Desgn Parameter selecton: Wˆ 0 = 0 Q P= E = I [formula gven] = ε λ P max 3 ( ε s unknown ry wth teraton) Dr. Radhakant Padh, AE Dept., IISc-Bangalore 7

Neuro-adaptve Desgn: Implementaton of Controller () Weght update Rule: ˆ W = γ e Pbβ, f e > E 0 x 0 where, e =, b=, x 1 β,, : Bass functon selecton P = ( X X U) γ : Learng rate otherwse ( U: pseudo control ) Dr. Radhakant Padh, AE Dept., IISc-Bangalore 8

Neuro-adaptve Desgn: Implementaton of Controller (3) Control Computaton: Adaptve Control: Pseudo Control: Actual Control: Uˆ = Wˆ U = U ˆ I Uad = X + K X + K X Uˆ ˆ δ ad = = fˆ 1 β c d p ad ( X, X, U) NN,, 1 ( X X U) Dr. Radhakant Padh, AE Dept., IISc-Bangalore 9

Neuro-adaptve desgn Nce Results: () Adapton happens n fnte tme () () () As t, the error e t les nsde the deadzone and W a constant value. t approaches Dr. Radhakant Padh, AE Dept., IISc-Bangalore 30

Promsng Results: (Flght control Problem) Dr. Radhakant Padh, AE Dept., IISc-Bangalore 31

References B. S. Km and A. J. Calse, Nonlnear Flght Control Usng Neural Networks, Journal of Gudance, Control and Dynamcs, Vol. 0, No. 1, 1997, pp. 6-33. J. Letner, A. J. Calse and J. V. R. Prasad, Analyss of Adaptve Neural Networks for Helcopter Flght Controls, Journal of Gudance, Control, and Dynamcs, Vol. 0, No. 5, 1997, pp. 97-979. M. Sharma and A. J. Calse, Neural-Network Augmentaton of Exstng Lnear Controllers, J. of Gudance, Control and Dynamcs, Vol. 8, No. 1, 005, pp. 1-19. Dr. Radhakant Padh, AE Dept., IISc-Bangalore 3

Dr. Radhakant Padh, AE Dept., IISc-Bangalore 33