Fine Local Tuning

You might also like

Download as pptx, pdf, or txt
Download as pptx, pdf, or txt
You are on page 1of 22

FINE LOCAL TUNING

PREPARED BY: GEK CAGATAN

FINE LOCAL TUNING


Problems in LOCAL SEARCHING arises when:
The DOMAINS of parameters are UNLIMITED The NUMBER of parameters are quite LARGE

HIGH PRECISION is REQUIRED


The performance of GAs is quite poor

FINE LOCAL TUNING


To IMPROVE Fine Local Tuning capabilities of GAs for HIGH

PRECISION problems,
We use SPECIAL MUTATION OPERATOR (with performance

quite different from traditional*)


TRADITIONAL
only one chromosome is changed at a time uses only local knowledge- only the bit undergoing mutation is known ABCDEFGH

TEST CASES

The task of designing & implementing algorithms for the solution of OPTIMAL

CONTROL PROBLEMS is a DIFFICULT ONE


Algorithm breaks down on problems of MODERATE SIZE & COMPLEXITY Difficult to deal with numerically DYNAMIC OPTIMIZATION SPECIFIC METHODS (DOSM) can be used. If used, they will

even be more difficult for a layman to handle.

OPTIMAL CONTROL PROBLEMS


GENETIC ALGORITHMS are used in this application.
GAs are modified to enhance its performance.

We show the quality & applicability of the developed system by a comparative study of some

dynamic optimization problems

There are 3 simple Discrete-Time Optimal Control Models (these are frequently used in the

applications of optimal control):


LINEAR-QUADRATIC PROBLEM HARVEST PROBLEM PUSH CART PROBLEM

General Algebraic Modelling System (GAMS) with MINOS Optimizer- a standard computational package was

used for solving such problems.

LINEAR- QUADRATIC PROBLEM

HARVEST PROBLEM

PUSH CART PROBLEM

THE REPRESENTATIONS

In floating point representation each chromosome vector is coded as a vector

of floating point numbers of the same length as the solution vector.


The PRECISION depends on the underlying MACHINE (but generally much

better than that of the binary representation).


We can EXTEND the precision of binary representation but it will SLOW DOWN the

algorithm.
FLOATING POINT representation is capable of representing quite LARGE DOMAINS.

THE SPECIALIZED OPERATORS


MUTATION and the CROSSOVER group

MUTATION GROUP

CROSSOVER GROUP

EXPERIMENTS & RESULTS


We present the results of the evolution program for the optimal control problems. POP. SIZE= 70 (FIXED) RUNS were made for 40,000 generations

LINEAR- QUADRATIC PROBLEM

EXPERIMENTS & RESULTS

HARVEST PROBLEM

PUSH CART PROBLEM

EVOLUTION PROGRAM VERSUS OTHER METHODS

EVOLUTION PROGRAM VERSUS OTHER METHODS

EVOLUTION PROGRAM VERSUS OTHER METHODS

EVOLUTION PROGRAM VERSUS OTHER METHODS PUSH CART PROBLEM

THE SIGNIFICANCE OF NON-UNIFORM MUTATION

THE SIGNIFICANCE OF NON-UNIFORM MUTATION

CONCLUSION

The NON-UNIFORM mutation improved the fine local tuning capabilities of the GA.

CONCLUSION
The NUMERICAL RESULTS were compared with those obtained from a search- based

computational package (GAMS).


While the Evolution Program gave us results comparable with the analytic solutions for all test

problems, GAMS failed for one of them.


The developed evolution program displayed some qualities not always present in the other

systems:
Optimization factor for the evolution program need not be continuous. Optimization packages are all-or-nothing propositions. Evolution programs give the users additional flexibility. The user can specify the computation time. The improvement of the performance of the system is often difficult for other optimization techniques.

You might also like