0% found this document useful (0 votes)

192 views

Dynamic Optimization

Dynamic optimization refers to minimizing or maximizing an objective function over time, as opposed to static optimization which considers only a single instant in time. Dynamic optimization has applications in human movement modeling as it allows investigation of optimal motor control over a period of time. Common solution methods include calculus of variations, optimal control theory, and approximating dynamic problems with static optimization techniques. Performance is typically evaluated by comparing model predictions to experimental measures like EMG or movement outcomes.

Uploaded by

Anurag Verma

Available Formats

Download as PDF, TXT or read online on Scribd

0% found this document useful (0 votes)

192 views

Dynamic Optimization

Uploaded by

Anurag Verma

Available Formats

Download as PDF, TXT or read online on Scribd

You are on page 1/ 17

8A-1

DYNAMIC OPTIMIZATION: BASICS

Lecture Overview
Static versus dynamic optimization? What applications? Why should optimization apply in human movement? How can dynamic optimization be implemented? What objective functions are relevant? How are dynamic optimization problems solved? Solution procedures - Calculus of variation Solution procedures - Optimal control

Solution procedures - Static optimization

Evaluation criteria Generalized performance criterion

John H. Challis - Modeling in Biomechanics

8A-2

STATIC VERSUS DYNAMIC OPTIMIZATION?

Optimization is the process of minimizing or
maximizing the costs/benefits of some action.

Static Optimization refers to the process of

minimizing or maximizing the costs/benefits of some objective function for one instant in time only.

Dynamic Optimization refers to the process of

minimizing or maximizing the costs/benefits of some objective function over a period of time. Sometimes called optimal control. In both cases equality and inequality constraints can be enforced.

John H. Challis - Modeling in Biomechanics

8A-3

WHAT APPLICATIONS?
Optimization has many applications: Economics Mechanical Designs Geographical Models Providing models with a nervous system Jumping FES Locomotion Reaching tasks and many others

John H. Challis - Modeling in Biomechanics

8A-4

WHY SHOULD OPTIMIZATION APPLY IN HUMAN MOVEMENT?

(Repeat)

Examples of Optimized Movement

Cotes and Meade (1960) examining horizontal treadmill walking showed that the stride frequency selected by the walkers required the least energy expenditure, compared with walking at stride frequencies less than or greater than the selected frequency. Cavanagh and Williams (1982) examining running at a constant speed showed subjects selected a stride length which minimized the energy cost for running at that speed. McMahon, Valiant, and Frederick (1987) presented evidence to suggest that the preferred style of running also requires the lowest energy cost. Brett (1965) presented evidence to suggest that swimming salmon also select the most efficient style. Consider natural selection.
John H. Challis - Modeling in Biomechanics

8A-5

HOW CAN DYNAMIC OPTIMIZATION BE IMPLEMENTED?

There are four components of optimization process:1. A dynamic system to be controlled 2. A performance criterion which defines movement goal 3. A controller which acts on the system 4. An algorithm which implements the control

John H. Challis - Modeling in Biomechanics

8A-6

HOW CAN DYNAMIC OPTIMIZATION BE IMPLEMENTED?

The task is the following: produce a movement with specified beginning and end positions which minimizes the muscular stresses squared. Objective Function
Minimize
t end n =7

t start

Fi i =1 CSAi

Inequality constraints j ( t end ) j/ ( t end ) errortol

j ( t end ) j/ ( t end ) errortol

Where
n number of muscles in the model Fi moment produced by the ith muscle CSAi - ith muscle cross-sectional area j , j joint angles and angular velocities

/ j , / j joint angles and angular velocities predicted by the model errortol - measure of the error in the data, constrains how close final model joint data should be to the measured data.
John H. Challis - Modeling in Biomechanics

8A-7

WHAT OBJECTIVE FUNCTIONS ARE RELEVANT?

Equivalent Question: How does CNS control
human movement? Jerk (e.g. arm movements)
2 + y 2 dt Minimize x
t start t end

Muscle Stress (e.g. walking)

Minimize
t end NM

t start

Fi i =1 CSAi

Mechanical Energy (e.g. Jumping)

Maximise J ,, t f = YCM t f +

( )

2 CM Y tf

( )

2. g

There are many others.

John H. Challis - Modeling in Biomechanics

8A-8

HOW ARE DYNAMIC OPTIMIZATION PROBLEMS SOLVED?

There are a variety of ways in which dynamic optimization problems can be solved and it depends on the nature of the objective function.

Calculus of Variation (e.g. Akhiezer, 1962)

Permits some problems to be simplified, others can be solved analytically

Optimal Control (e.g. Polak and Mayne, 1975)

There are software packages which allow the solution of optimal control problems.

Static Optimization (e.g. Goh and Teo, 1988)

It is sometimes possible to use static optimization techniques to solve dynamic optimization problems.

John H. Challis - Modeling in Biomechanics

8A-9

SOLUTION PROCEDURES
Calculus of Variation
[Permits some problems to be simplified, others can be solved analytically.] Consider the problem of moving the load shown schematically

The resistance to motion is a function of the mass (M), the degree of damping (D), and the stiffness (K). The task is to move the mass from x=0 to x=X in time T.

John H. Challis - Modeling in Biomechanics

8A-10

SOLUTION PROCEDURES
Calculus of Variation
If the equations of motion of the system are written and the objective function formulated an analytical solution can be obtained.
2 )dt Minimize Velocity Minimize (x
T

Gives

= c1 x x = c1 .t + co

Where c1 , co are constants depndent on intitial and ifnal conditions.

Minimize Acceleration
= 2.c2 + 6.c3 .t x

2 dt Minimize x
o

( )

Gives

= 2.c2 .t + 3.c3 .t 2 x x = c2 .t 2 + c3 .t 3

Where c0 , c1 , c2 , c3 , are constants dependent on intitial and final conditions.

John H. Challis - Modeling in Biomechanics

8A-11

SOLUTION PROCEDURES
Calculus of Variation Minimize Jerk
Gives
5 i =0

2 dt Minimize x
o

( )

x = c i .t i

Similar simple equations are available for the analytical solution to movement problems commonly accounted in human movement (e.g. Flash and Hogan, 1985).

John H. Challis - Modeling in Biomechanics

8A-12

SOLUTION PROCEDURES
Optimal Control
There are software packages which allow the solution of optimal control problems. The theory behind these techniques is similar to those for static optimization (Polak and Mayne, 1975; Teo and Moore, 1978). The objective function is often simplified using calculus of variation. (see Bryson and Ho, 1975). These techniques are generally time consuming, for example Anderson et al. (1995) estimated to simulate a vertical jump using a four segment, eight muscle model took three days of CPU time. Consider an 8 segment, 14 degree of freedom model, actuated by 46 muscle models comprising 120 first-order differential equations. One iteration of an optimal control algorithm takes the following times IRIS 21.36 hours Cray 0.77 Hours 128 Parallel Processors 0.88 hours It was estimated that 100 iterations would be required for convergence.

John H. Challis - Modeling in Biomechanics

8A-13

SOLUTION PROCEDURES
Static Optimization
Goh and Teo, (1988) proposed solving optimal control problems using static optimization algorithms. The optimal control problem is converted into a parameter optimization approach by dividing up the control history (e.g. muscle activations) into a series of knots from which the full time history can be determined by interpolation. This approach normally gives equivalent results to traditional optimal control solutions, but the algorithms used to implement these procedures are generally numerically more robust. Another advantage is that the solutions are normally attained using less processing time (Goh and Teo, 1988).

John H. Challis - Modeling in Biomechanics

8A-14

SOLUTION PROCEDURES
Static Optimization
Therefore the static optimization techniques can be used, for example Direct Search Technique - Hooke and Jeeves (1961) Downhill Simplex - Nelder and Mead, (1965) Gradient Search - Powell, 1970). This approach is particularly well suited to problems which assume a singular bang-bang control for the muscles.

John H. Challis - Modeling in Biomechanics

8A-15

EVALUATION CRITERIA
Two Principle Methods
EMG - Use EMG signal to indicate when the muscles compare this with the activity of the muscles as predicted by the optimization procedure. Performance Measures does the predicted movement correspond with actual movement, and/or is the movement outcome the same as with real subjects (e.g. movement time, jump height).

John H. Challis - Modeling in Biomechanics

8A-16

GENERALIZED PERFORMANCE CRITERION

Zajac and Winters (1990) proposed the following objective function which they termed a generalized performance criterion
Minimize J = J TK + J NM + J BJ

Where the three components are Task specific Kinematics track a given trajectory, make hand follow straight line, minimize jerk, etc.

NeuroMuscular minimize muscle stresses, neural

effort, etc.

Bone Joint minimize contact forces, avoid certain

ranges of motion, etc.

John H. Challis - Modeling in Biomechanics

8A-17

REVIEW QUESTIONS
1) What is meant by dynamic optimization? 2) Why is dynamic optimization considered more physiologically realistic than static optimization? 3) Give some examples of objective functions which can be used for dynamic optimization (name source of function). What are their application? 4) How can dynamic optimization problems be solved? 5) What are the evaluation criteria for dynamic optimization problems? 6) Present a generalized performance criteria that you think could be applicable to some aspect of human movement (name the movement).