Placement by Thermodynamic Simulated Annealing

18
7/21/2019 Placement by Thermodynamic Simulated Annealing http://slidepdf.com/reader/full/placement-by-thermodynamic-simulated-annealing 1/18 Placement by Thermodynamic Simulated Annealing CSE624 Heuristic Optimization Mahmud Rasih ÇELENLİOĞLU October 2013

description

Placement by ThermodynamicSimulated Annealing

Transcript of Placement by Thermodynamic Simulated Annealing

Page 1: Placement by Thermodynamic Simulated Annealing

7/21/2019 Placement by Thermodynamic Simulated Annealing

http://slidepdf.com/reader/full/placement-by-thermodynamic-simulated-annealing 1/18

Placement by Thermodynamic

Simulated AnnealingCSE624 Heuristic Optimization

Mahmud Rasih ÇELENLİOĞLU

October 2013

Page 2: Placement by Thermodynamic Simulated Annealing

7/21/2019 Placement by Thermodynamic Simulated Annealing

http://slidepdf.com/reader/full/placement-by-thermodynamic-simulated-annealing 2/18

Outline

2

• Placement Problem

• Contribution of Paper

Proposed Method

• Results

• Conclusion

Page 3: Placement by Thermodynamic Simulated Annealing

7/21/2019 Placement by Thermodynamic Simulated Annealing

http://slidepdf.com/reader/full/placement-by-thermodynamic-simulated-annealing 3/18

Page 4: Placement by Thermodynamic Simulated Annealing

7/21/2019 Placement by Thermodynamic Simulated Annealing

http://slidepdf.com/reader/full/placement-by-thermodynamic-simulated-annealing 4/18

Integrated Circuit Placement Problem

4

The placement quality affects both the area and

speed of circuits.

Page 5: Placement by Thermodynamic Simulated Annealing

7/21/2019 Placement by Thermodynamic Simulated Annealing

http://slidepdf.com/reader/full/placement-by-thermodynamic-simulated-annealing 5/18

Contribution of the Paper

5

• In combinatorial optimization problems, SA issuccessfull.

• However, SA requires costly experimental studies infine tuning the most suitable annealing schedule.

• A new schedule is derived from thermodynamiclaws.

• Temperature in this method is free to evolve and itsvalue is continuously updated from the variation ofstate functions as the internal energy and entropy.

Page 6: Placement by Thermodynamic Simulated Annealing

7/21/2019 Placement by Thermodynamic Simulated Annealing

http://slidepdf.com/reader/full/placement-by-thermodynamic-simulated-annealing 6/18

Parameters & Functions

6

• T0 = -avg(abs( Δ C))/ ln(P) & Tend = 0.0025

ΔC: cost variationLn(P): high probability of acceptance

• How to find temperature T?

Derived from 1. & 2. thermodynamic laws Rule 1: Conservation of total energy Rule 2: Efficiency is always lower than %100

Cost Evaluation Function Bounding Box

• What is stopping condition?

T < Tend && T is stabilized

Page 7: Placement by Thermodynamic Simulated Annealing

7/21/2019 Placement by Thermodynamic Simulated Annealing

http://slidepdf.com/reader/full/placement-by-thermodynamic-simulated-annealing 7/18

Bounding Box

7

span of BB in direction X & Y

average channel capacity in track X, Y

(constant)compensates underestimated area of multiterminal nets

(depends on the # of terminals of net n)(n <= 3 then q=1, n = 50 then q=2.79 )

Page 8: Placement by Thermodynamic Simulated Annealing

7/21/2019 Placement by Thermodynamic Simulated Annealing

http://slidepdf.com/reader/full/placement-by-thermodynamic-simulated-annealing 8/18

Probability of Accepting New Solution

8

• In physics, the Boltzmann factor is a weighting factor that determines the relative probability of a particle to be in a state in a multi-state system in thermodynamic equilibrium (there are no netflows of matter or energy) at temperature T

• ΔC = C B - C A

accept if cost is decreased

Boltzman Factor

Page 9: Placement by Thermodynamic Simulated Annealing

7/21/2019 Placement by Thermodynamic Simulated Annealing

http://slidepdf.com/reader/full/placement-by-thermodynamic-simulated-annealing 9/18

Thermodynamic Simulated Annealing

9

• It provides an alternative method to performcooling close to equilibrium without experimentaladjustments

• TSA is a reversible process that is intermediate statesare also equilibrium states

• Temperature is free to evolve (not controlled) butcan be measured from thermodynamic laws

cost variation

entropy variation

Page 10: Placement by Thermodynamic Simulated Annealing

7/21/2019 Placement by Thermodynamic Simulated Annealing

http://slidepdf.com/reader/full/placement-by-thermodynamic-simulated-annealing 10/18

Thermodynamic Simulated Annealing

10

• In information theory, Shannon entropy quantifiesthe expected value of the information contained ina message (ΔI = -ln(P) )

• Δ I = -ln(P) is converted to Δ S = ln(P)

set of transformations accepteduntil kth iteration set of movements tried withincrement in cost until kth iteration

run-time/quality trade off

Page 11: Placement by Thermodynamic Simulated Annealing

7/21/2019 Placement by Thermodynamic Simulated Annealing

http://slidepdf.com/reader/full/placement-by-thermodynamic-simulated-annealing 11/18

• Negative values appear in initial stages whenC i > C 0. To prevent this;

Thermodynamic Simulated Annealing

11

increment in costno change in entropy

Page 12: Placement by Thermodynamic Simulated Annealing

7/21/2019 Placement by Thermodynamic Simulated Annealing

http://slidepdf.com/reader/full/placement-by-thermodynamic-simulated-annealing 12/18TSA Algorithm 12

--swap 2 blocks or move--block to an empty position--Bounding Box

--Boltzmann Factor

Page 13: Placement by Thermodynamic Simulated Annealing

7/21/2019 Placement by Thermodynamic Simulated Annealing

http://slidepdf.com/reader/full/placement-by-thermodynamic-simulated-annealing 13/18

• Range limiter: Maximum distance that a block canbe moved is linearly shortened with temperature

• It is calculated once the 80% of the acceptancerate is reached

Performance Improvement

13

Page 14: Placement by Thermodynamic Simulated Annealing

7/21/2019 Placement by Thermodynamic Simulated Annealing

http://slidepdf.com/reader/full/placement-by-thermodynamic-simulated-annealing 14/18

14

Results

Page 15: Placement by Thermodynamic Simulated Annealing

7/21/2019 Placement by Thermodynamic Simulated Annealing

http://slidepdf.com/reader/full/placement-by-thermodynamic-simulated-annealing 15/18

15

Results

Page 16: Placement by Thermodynamic Simulated Annealing

7/21/2019 Placement by Thermodynamic Simulated Annealing

http://slidepdf.com/reader/full/placement-by-thermodynamic-simulated-annealing 16/18

16

Results

Page 17: Placement by Thermodynamic Simulated Annealing

7/21/2019 Placement by Thermodynamic Simulated Annealing

http://slidepdf.com/reader/full/placement-by-thermodynamic-simulated-annealing 17/18

17

Conclusion

• TSA provides a different cooling schedule derivedfrom thermodynamic laws

• Temperature drop is not controlled externally butcomputed from current state

Page 18: Placement by Thermodynamic Simulated Annealing

7/21/2019 Placement by Thermodynamic Simulated Annealing

http://slidepdf.com/reader/full/placement-by-thermodynamic-simulated-annealing 18/18

18

Thank You

&

Questions ?