evolving digital hardware ehw module 200825 tone discriminator results a successful individual...
TRANSCRIPT
![Page 1: Evolving Digital Hardware EHW Module 200825 Tone Discriminator Results A successful individual (after 5000 generations). The left hand image shows the cell connectivity for the full](https://reader034.vdocuments.site/reader034/viewer/2022042122/5e9cca081d56b14b596dbfd8/html5/thumbnails/1.jpg)
1
Evolving Digital HardwareEvolving Digital HardwareEHW Module 2008EHW Module 2008
Andy GreenstedDepartment of [email protected]
Lecture 3
www.bioinspired.com/users/ajg112/teaching/evoHW.shtml
![Page 2: Evolving Digital Hardware EHW Module 200825 Tone Discriminator Results A successful individual (after 5000 generations). The left hand image shows the cell connectivity for the full](https://reader034.vdocuments.site/reader034/viewer/2022042122/5e9cca081d56b14b596dbfd8/html5/thumbnails/2.jpg)
2
CGPCGP(Cartesian Genetic Programming)(Cartesian Genetic Programming)
![Page 3: Evolving Digital Hardware EHW Module 200825 Tone Discriminator Results A successful individual (after 5000 generations). The left hand image shows the cell connectivity for the full](https://reader034.vdocuments.site/reader034/viewer/2022042122/5e9cca081d56b14b596dbfd8/html5/thumbnails/3.jpg)
3
CGPCGP● Form of GP based on Acyclic directed graphs
– Re-use of nodes in graph.
● Fixed genotype length
– List of integers encoding nodes and connection in graph.
● Bounded variable length phenotype
– Not all nodes are connected.
Cartesian Genetic Programming. J. Miller and P. Thomson,Proceedings of the 3rd European Conference on Genetic Programming, Edinburgh, (2000) 121-132.
![Page 4: Evolving Digital Hardware EHW Module 200825 Tone Discriminator Results A successful individual (after 5000 generations). The left hand image shows the cell connectivity for the full](https://reader034.vdocuments.site/reader034/viewer/2022042122/5e9cca081d56b14b596dbfd8/html5/thumbnails/4.jpg)
4
CGP Evolution StrategyCGP Evolution Strategy● CGP uses a (1+4) evolution strategy
● Generations are created from:
– The current fittest, unless a new equally fit or fitter solution has been found.
– Mutations of the fittest from the previous generation.
Fitter individuals have lower fitness score
![Page 5: Evolving Digital Hardware EHW Module 200825 Tone Discriminator Results A successful individual (after 5000 generations). The left hand image shows the cell connectivity for the full](https://reader034.vdocuments.site/reader034/viewer/2022042122/5e9cca081d56b14b596dbfd8/html5/thumbnails/5.jpg)
5
Example Circuit in CGP RepresentationExample Circuit in CGP Representation
![Page 6: Evolving Digital Hardware EHW Module 200825 Tone Discriminator Results A successful individual (after 5000 generations). The left hand image shows the cell connectivity for the full](https://reader034.vdocuments.site/reader034/viewer/2022042122/5e9cca081d56b14b596dbfd8/html5/thumbnails/6.jpg)
6
CGP Encoding FeaturesCGP Encoding Features● The genotype is a fixed length
● Function inputs can only connect to previous function outputs
– Stops combinatorial loops
![Page 7: Evolving Digital Hardware EHW Module 200825 Tone Discriminator Results A successful individual (after 5000 generations). The left hand image shows the cell connectivity for the full](https://reader034.vdocuments.site/reader034/viewer/2022042122/5e9cca081d56b14b596dbfd8/html5/thumbnails/7.jpg)
7
Case Study 1Case Study 1Evolving Fault ToleranceEvolving Fault Tolerance
![Page 8: Evolving Digital Hardware EHW Module 200825 Tone Discriminator Results A successful individual (after 5000 generations). The left hand image shows the cell connectivity for the full](https://reader034.vdocuments.site/reader034/viewer/2022042122/5e9cca081d56b14b596dbfd8/html5/thumbnails/8.jpg)
8
Evolved Fault ToleranceEvolved Fault Tolerance
● Use evolution to create fault-tolerant circuits.● A circuit's fitness is based upon:
– Its ability to operate correctly under different fault conditions.
Evolution of Fault-Tolerant and Noise-Robust Digital Designs M.Hartmann, P.C. Haddow, IEE Proc.-Comput. Digit. Tech. Vol 151, No.4 , July 2004
F Average fitness for all environmentsCm A circuit under environment mT The target circuitdiff() The number of different outputs bitsTPI Number of Environments
![Page 9: Evolving Digital Hardware EHW Module 200825 Tone Discriminator Results A successful individual (after 5000 generations). The left hand image shows the cell connectivity for the full](https://reader034.vdocuments.site/reader034/viewer/2022042122/5e9cca081d56b14b596dbfd8/html5/thumbnails/9.jpg)
9
The Gate ModelThe Gate Model
E1: Input Error
E2: Input Error
F: Function Generator
E3: Output Error
N: Noise
Errors: Stuck-at errorsFloating outputPartially random output
Functions: NOR, NAND, OR, AND, NOT, VCC & GND
![Page 10: Evolving Digital Hardware EHW Module 200825 Tone Discriminator Results A successful individual (after 5000 generations). The left hand image shows the cell connectivity for the full](https://reader034.vdocuments.site/reader034/viewer/2022042122/5e9cca081d56b14b596dbfd8/html5/thumbnails/10.jpg)
10
The Gate ModelThe Gate Model● The model for a NOR gate
InA InB NOR
0 0 10 1 01 0 01 1 0
(0,1)
(0,0)
(1,0)
(1,1)
![Page 11: Evolving Digital Hardware EHW Module 200825 Tone Discriminator Results A successful individual (after 5000 generations). The left hand image shows the cell connectivity for the full](https://reader034.vdocuments.site/reader034/viewer/2022042122/5e9cca081d56b14b596dbfd8/html5/thumbnails/11.jpg)
11
Target CircuitsTarget Circuits● Both 2bit Adders and Multipliers were tested
![Page 12: Evolving Digital Hardware EHW Module 200825 Tone Discriminator Results A successful individual (after 5000 generations). The left hand image shows the cell connectivity for the full](https://reader034.vdocuments.site/reader034/viewer/2022042122/5e9cca081d56b14b596dbfd8/html5/thumbnails/12.jpg)
12
ResultsResults
● Average fitness of circuits against applied noise level.
● Average number of used gates against noise applied during evolution.
● (Evolution prefers circuits with less gates as noise level increases, this reduces the likelihood of error)
![Page 13: Evolving Digital Hardware EHW Module 200825 Tone Discriminator Results A successful individual (after 5000 generations). The left hand image shows the cell connectivity for the full](https://reader034.vdocuments.site/reader034/viewer/2022042122/5e9cca081d56b14b596dbfd8/html5/thumbnails/13.jpg)
13
Case Study 2Case Study 2RISARISA
![Page 14: Evolving Digital Hardware EHW Module 200825 Tone Discriminator Results A successful individual (after 5000 generations). The left hand image shows the cell connectivity for the full](https://reader034.vdocuments.site/reader034/viewer/2022042122/5e9cca081d56b14b596dbfd8/html5/thumbnails/14.jpg)
14
The Reconfigurable Integrated System The Reconfigurable Integrated System Array (RISA) ArchitectureArray (RISA) Architecture
● A new embryonic tissue
– Processor/FPGA Array
● Supports Different Configurations:
– Processor Array
● Systolic Arrays– FPGA
● Normal FPGA Applications– Processor/FPGA Array
● Intrinsic reconfiguration● Seamless creation of larger array
by connecting devices
![Page 15: Evolving Digital Hardware EHW Module 200825 Tone Discriminator Results A successful individual (after 5000 generations). The left hand image shows the cell connectivity for the full](https://reader034.vdocuments.site/reader034/viewer/2022042122/5e9cca081d56b14b596dbfd8/html5/thumbnails/15.jpg)
15
RISA FPGA Fabric RISA FPGA Fabric
● The RISA FPGA Fabric provides a platform for combinatorial circuit evolution
![Page 16: Evolving Digital Hardware EHW Module 200825 Tone Discriminator Results A successful individual (after 5000 generations). The left hand image shows the cell connectivity for the full](https://reader034.vdocuments.site/reader034/viewer/2022042122/5e9cca081d56b14b596dbfd8/html5/thumbnails/16.jpg)
16
RISA FPGA Fabric - Function UnitRISA FPGA Fabric - Function Unit
● Function Generator
– 4 input LUT
– 4x1 RAM Block
– Variable length Shift Register
● Full Adder Gates
● Dedicated Multiplexor
● D Flip Flop
● Carry Chain
● Shift Chain
![Page 17: Evolving Digital Hardware EHW Module 200825 Tone Discriminator Results A successful individual (after 5000 generations). The left hand image shows the cell connectivity for the full](https://reader034.vdocuments.site/reader034/viewer/2022042122/5e9cca081d56b14b596dbfd8/html5/thumbnails/17.jpg)
17
FabricationFabrication● Each device contains 1 RISA
cell
– 6x6 Cluster FPGA Fabric
– 6 IO Blocks per side
– 1 SNAP core
– 16kB Memory (8192 16b words)
● 180nm process
● 5x5mm die area
● Cell based ASIC
![Page 18: Evolving Digital Hardware EHW Module 200825 Tone Discriminator Results A successful individual (after 5000 generations). The left hand image shows the cell connectivity for the full](https://reader034.vdocuments.site/reader034/viewer/2022042122/5e9cca081d56b14b596dbfd8/html5/thumbnails/18.jpg)
18
First Experiments using theFirst Experiments using theRISA PlatformRISA Platform
● Xilinx Spartan-3E device connected under RISA device
– Able to apply test vectors to RISA device
– The spartan controls RISA configuration
![Page 19: Evolving Digital Hardware EHW Module 200825 Tone Discriminator Results A successful individual (after 5000 generations). The left hand image shows the cell connectivity for the full](https://reader034.vdocuments.site/reader034/viewer/2022042122/5e9cca081d56b14b596dbfd8/html5/thumbnails/19.jpg)
19
Evolving a simple digital circuitEvolving a simple digital circuit● A simple evolvable hardware experiment for testing the device
– Evolution of a 4 bit parity generator (Both even and odd parity)
● Evolutionary algorithm runs on a Microblaze core within the Spartan FPGA
● Candidate solutions loaded into RISA FPGA fabric and test vectors applied through RISA IO Blocks
![Page 20: Evolving Digital Hardware EHW Module 200825 Tone Discriminator Results A successful individual (after 5000 generations). The left hand image shows the cell connectivity for the full](https://reader034.vdocuments.site/reader034/viewer/2022042122/5e9cca081d56b14b596dbfd8/html5/thumbnails/20.jpg)
20
Fitness CurvesFitness Curves
● Population Size: 32
● Tournament size: 4
● Mutation rate: (1,2,4)/256
● Elitism
Extrinsic Evolvable Hardware on the RISA Architecture, Andrew Greensted and Andy Tyrrell, ICES 2007, LNCS Evolvable Systems: From Biology to Hardware, 4684:244-255, Wuhan, China, September 2007
![Page 21: Evolving Digital Hardware EHW Module 200825 Tone Discriminator Results A successful individual (after 5000 generations). The left hand image shows the cell connectivity for the full](https://reader034.vdocuments.site/reader034/viewer/2022042122/5e9cca081d56b14b596dbfd8/html5/thumbnails/21.jpg)
21
Case Study 3Case Study 3Tone DiscriminatorTone Discriminator
![Page 22: Evolving Digital Hardware EHW Module 200825 Tone Discriminator Results A successful individual (after 5000 generations). The left hand image shows the cell connectivity for the full](https://reader034.vdocuments.site/reader034/viewer/2022042122/5e9cca081d56b14b596dbfd8/html5/thumbnails/22.jpg)
22
Thompson's Tone DiscriminatorThompson's Tone Discriminator
● Adrian Thompson (Sussex University)
● The aim was to evolve a digital circuit that could discriminate between two frequencies of a signal input to the system.
● The experiment was conducted in 1996
An Evolved Circuit, Intrinsic in Silicon, Entwined with Physics. Adrian Thompson, 1st International Conference on Evolvable Systems 1996, (c) Springer Verlag 1997
![Page 23: Evolving Digital Hardware EHW Module 200825 Tone Discriminator Results A successful individual (after 5000 generations). The left hand image shows the cell connectivity for the full](https://reader034.vdocuments.site/reader034/viewer/2022042122/5e9cca081d56b14b596dbfd8/html5/thumbnails/23.jpg)
23
Thompson's Tone DiscriminatorThompson's Tone Discriminator
A Xilinx XC6216 was used to evolve the tone discriminator.● No clock input signal was used.
● Therefore there was no timing reference● The evolved circuit had to discriminate between a 1KHz and
10KHz input signal.● The area used by evolution was limited to a 10x10 array of cells
in one corner. 100 out of the device's 4096 cells.
![Page 24: Evolving Digital Hardware EHW Module 200825 Tone Discriminator Results A successful individual (after 5000 generations). The left hand image shows the cell connectivity for the full](https://reader034.vdocuments.site/reader034/viewer/2022042122/5e9cca081d56b14b596dbfd8/html5/thumbnails/24.jpg)
24
Tone Discriminator ResultsTone Discriminator Results
● Signal propagation time in the XC6216 technology is in the order of nanoseconds.
● The circuit was still able to discriminate between signals with periods in the order of milliseconds
![Page 25: Evolving Digital Hardware EHW Module 200825 Tone Discriminator Results A successful individual (after 5000 generations). The left hand image shows the cell connectivity for the full](https://reader034.vdocuments.site/reader034/viewer/2022042122/5e9cca081d56b14b596dbfd8/html5/thumbnails/25.jpg)
25
Tone Discriminator ResultsTone Discriminator ResultsA successful individual (after 5000 generations).
● The left hand image shows the cell connectivity for the full 10x10 array.● The right hand image shows the cells required for correct operation.● The cells shaded grey contain no connected logic, but their
configuration is required for correct operation.● The solution was very device and environment dependent.
![Page 26: Evolving Digital Hardware EHW Module 200825 Tone Discriminator Results A successful individual (after 5000 generations). The left hand image shows the cell connectivity for the full](https://reader034.vdocuments.site/reader034/viewer/2022042122/5e9cca081d56b14b596dbfd8/html5/thumbnails/26.jpg)
26
Tone Discriminator ResultsTone Discriminator Results
(Low Freq)(High Freq)
IncreasingTemperature
![Page 27: Evolving Digital Hardware EHW Module 200825 Tone Discriminator Results A successful individual (after 5000 generations). The left hand image shows the cell connectivity for the full](https://reader034.vdocuments.site/reader034/viewer/2022042122/5e9cca081d56b14b596dbfd8/html5/thumbnails/27.jpg)
27
Case Study 4Case Study 4Prime Number GeneratorPrime Number Generator
![Page 28: Evolving Digital Hardware EHW Module 200825 Tone Discriminator Results A successful individual (after 5000 generations). The left hand image shows the cell connectivity for the full](https://reader034.vdocuments.site/reader034/viewer/2022042122/5e9cca081d56b14b596dbfd8/html5/thumbnails/28.jpg)
28
The CompetitionThe Competition● GECCO 2006 (Genetic and Evolutionary Computation Conference)
Consecutive Primes Competition
Evolve a polynomial with integer coefficients such that given an integer value i as input produces the ith prime number, p(i), for the largest possible value of i.
So, if f(i) is the evolved function, we expect:f(1)=2,f(2)=3,f(3)=5,f(4)=7,f(5)=11, etc...
![Page 29: Evolving Digital Hardware EHW Module 200825 Tone Discriminator Results A successful individual (after 5000 generations). The left hand image shows the cell connectivity for the full](https://reader034.vdocuments.site/reader034/viewer/2022042122/5e9cca081d56b14b596dbfd8/html5/thumbnails/29.jpg)
29
Prime Generating FunctionsPrime Generating Functions
Y
N
Y
N
Y
AllPositive
Y
N
Y
Y
Y
All Integer coefficients?
23
57
41
45
40
Len
44546738095860i + 56211383760397
(i5 - 133i4 + 6729i3 - 158379i2
+ 1720294i - 6823316)/4
i5 - 61i4 + 1339i3 - 12523i2 + 42398i + 11699
36i2 – 810i + 2753
i2 + i + 41
Polynomial
Frind, Jobling,
Underwood
Sunder Gupta
Wroblewski, Meyrignac
Ruby, Fung
Euler
Discoverer
The best so far...
![Page 30: Evolving Digital Hardware EHW Module 200825 Tone Discriminator Results A successful individual (after 5000 generations). The left hand image shows the cell connectivity for the full](https://reader034.vdocuments.site/reader034/viewer/2022042122/5e9cca081d56b14b596dbfd8/html5/thumbnails/30.jpg)
30
Adapting to CGPAdapting to CGPConvert polynomial to a boolean form:
p(i) = bm2m + bm-12m-1 + ... + b424 + b323 + b222 + b121 + b020
Each coefficient is a function of the binary values of the input prime number:
bm = fm(a0, a1, …, an)
...
b1 = f1(a0, a1, …, an)
b0 = f0(a0, a1, …, an)
● Multi-Chromosome CGP was used:● Essentially one instance of CGP per coefficient.
● Available operations:● AND NAND, OR, NOR, XOR, XNOR, AND (!in
0), OR (!in
0)
Where:a ∈ {0,1}b ∈ {0,1}
![Page 31: Evolving Digital Hardware EHW Module 200825 Tone Discriminator Results A successful individual (after 5000 generations). The left hand image shows the cell connectivity for the full](https://reader034.vdocuments.site/reader034/viewer/2022042122/5e9cca081d56b14b596dbfd8/html5/thumbnails/31.jpg)
31
16 Consecutive Primes16 Consecutive Primes
b5 = a0 + a1a2 – 2a0a1a2
b4 = -((a1(2a2 - 1) - a2)(1 + a2(a3a4 - 1))) + a0(1 - 2a2 + a22(2 - 2a3a4) + 2a1(2a2 - 1)(1 +
a2(a3a4 - 1)))
b3 = a2(a3 + a4 - 2a3a4) + a1a3(1 + a2(2a3a4 - a3 - a4))
b2 = 1 - a3 - a4 + 2a32a4 - 2a2
4(2a3 - 1)3(a4 - 1)a4 + 2a3a42 - 2a3
2a42 - a2
2(2a3 - 1)(a3(2 - 6a4) +
(3 - 2a4)a4 + 6a32(a4 - 1)a4) + a2
3(1 - 2a3)2(1 - (1 + 6a3)a4 + (6a3 -2)a42) + a2(2a3
3(a4 - 1)a4
+ 2a42 + a3(2 + 5a4 - 8a4
2) + a32(1 - 9a4 + 6a4
2) - 1) + a1(1 - a2a3 + a22(2a3 - 1))(2a3 + 2a4
-1 - a3a4 - 2a32a4 + 2a2
4(2a3 - 1)3(a4 - 1)a4 - 2a3a42 + 2a3
2a42 + 2a2
2(2a3 - 1)(a3(2 - 4a4) -
(a4 - 2)a4 + 3a32(a4 - 1)a4) - 2a2
3(1 - 2a3)2(1 - (1 + 3a3)a4 + (3a3 - 1)a42) - a2(a4 - 2 +
2a33(a4 - 1)a4 + 2a4
2 + a3(4 + 4a4 - 8a42) + 2a3
2(1 - 5a4 + 3a42)))
b1 = 1 - a3 + a32 - a2a3
2 - a1(a2(1 + a32 + a3(a4 - 3)) + a3
2(1 - 2a4) + a22a3(2a3 - 1)(a4 - 1)) -
a32a4 + a2a3
2a4 + a12(a2 - 1)a2a3(2a3 - 1)(2a4 - 1) - a0(2a1a2 - 1)(a3 - 1)(1 + (a2 - 1)a3(1 - a4
+ a1(2a4 - 1)))
b0 = a2 - a0(a1 - 1)(a2 - 1)(a3 - 1) + a1(a2 - 1)(a3 - 1) + a3 - a2a3
p(i) = b525 + b424 + b323 + b222 + b121 + b020 where
These equations have been simplified from their booleanoperation form.
![Page 32: Evolving Digital Hardware EHW Module 200825 Tone Discriminator Results A successful individual (after 5000 generations). The left hand image shows the cell connectivity for the full](https://reader034.vdocuments.site/reader034/viewer/2022042122/5e9cca081d56b14b596dbfd8/html5/thumbnails/32.jpg)
32
16 Consecutive Primes16 Consecutive Primes