1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
|
<chapter xmlns="http://docbook.org/ns/docbook"
xmlns:xlink="http://www.w3.org/1999/xlink"
xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance"
xsi:schemaLocation="http://docbook.org/ns/docbook http://www.docbook.org/xml/5.0/xsd/docbook.xsd"
id="watchmaker_chapter">
<title>The Watchmaker Framework</title>
<indexterm significance="preferred"><primary>Watchmaker Framework</primary></indexterm>
<para>
The Watchmaker Framework for Evolutionary Computation is an extensible, high-performance,
object-oriented framework for implementing platform-independent evolutionary algorithms
in Java.
It is freely available under a permissive Open Source licence. It can be downloaded from
<link xlink:href="http://watchmaker.uncommons.org">http://watchmaker.uncommons.org</link>.
</para>
<para>
This chapter introduces the core components of the Watchmaker Framework and shows how
they can be used to implement simple evolutionary algorithms such as the "Hello World"
example outlined in the previous chapter.
</para>
<section>
<title>The Evolution Engine</title>
<indexterm><primary><classname>GenerationalEvolutionEngine</classname></primary></indexterm>
<indexterm><primary><interfacename>EvolutionEngine</interfacename></primary></indexterm>
<para>
The central object of an evolutionary program built with the Watchmaker Framework is
the evolution engine.
</para>
<para>
The framework provides multiple implementations of the
<interfacename>EvolutionEngine</interfacename> interface, but the one that you will
usually want to use is <classname>GenerationalEvolutionEngine</classname>. This is a
general-purpose implementation of the evolutionary algorithm outline from chapter 1.
</para>
<para>
An <interfacename>EvolutionEngine</interfacename> has a single generic type parameter
that indicates the type of object that it can evolve.
For the "Hello World" program we need to be able to evolve Java strings.
Code that creates an engine that can evolve strings would look something like this:
</para>
<informalexample>
<programlisting language="java">
<![CDATA[EvolutionEngine<String> engine
= new GenerationalEvolutionEngine<String>(candidateFactory,
evolutionaryOperator,
fitnessEvaluator,
selectionStrategy,
rng);]]>
</programlisting>
</informalexample>
<para>
Once you have created an <interfacename>EvolutionEngine</interfacename>, your program
is as simple as calling the <methodname>evolve</methodname> method with appropriate
arguments.
However, as you can see from the code snippet above, there is a little bit of work to
be done first in order to create an <interfacename>EvolutionEngine</interfacename> that
is configured appropriately for the given problem.
The constructor of the <classname>GenerationalEvolutionEngine</classname> class requires
five objects. These are:
</para>
<itemizedlist>
<listitem>A Candidate Factory</listitem>
<listitem>An Evolutionary Operator</listitem>
<listitem>A Fitness Evaluator</listitem>
<listitem>A Selection Strategy</listitem>
<listitem>A Random Number Generator</listitem>
</itemizedlist>
</section>
<section>
<title>The Candidate Factory</title>
<indexterm significance="preferred"><primary><interfacename>CandidateFactory</interfacename></primary></indexterm>
<para>
The first object that needs to be plugged into the evolution engine is a candidate
factory. Every evolutionary simulation must start with an initial population of
candidate solutions and the <interfacename>CandidateFactory</interfacename> interface
is the mechanism by which the evolution engine creates this population.
</para>
<para>
A candidate factory implementation has an associated type. It can only create
objects of that type. The type must match the type of the evolution engine that
it is plugged into.
You can write your own implementation of <interfacename>CandidateFactory</interfacename>
for your program or, if you are using a common type such as strings, lists or
arrays, you may be able to use a ready-made factory from the
<package>org.uncommons.watchmaker.framework.factories</package> package.
</para>
<indexterm><primary><classname>StringFactory</classname></primary></indexterm>
<para>
For our "Hello World" program, we can use the provided
<classname>StringFactory</classname>:
</para>
<informalexample>
<programlisting language="java">
<![CDATA[// Define the set of permitted characters (A-Z plus space).
char[] chars = new char[27];
for (char c = 'A'; c <= 'Z'; c++)
{
chars[c - 'A'] = c;
}
chars[26] = ' ';
// Factory for random 11-character Strings.
CandidateFactory<String> factory = new StringFactory(chars, 11);]]>
</programlisting>
</informalexample>
<tip>
<indexterm significance="preferred"><primary><classname>AbstractCandidateFactory</classname></primary></indexterm>
<para>
When writing your own <interfacename>CandidateFactory</interfacename> implementations,
it is easiest to extend the provided <classname>AbstractCandidateFactory</classname>
base class since there is then only a single method that must be implemented.
</para>
</tip>
</section>
<section>
<title>Evolutionary Operators</title>
<indexterm significance="preferred"><primary><interfacename>EvolutionaryOperator</interfacename></primary></indexterm>
<para>
Evolutionary operators are the components that perform the actual evolution of a
population. Cross-over is an evolutionary operator, as is mutation.
</para>
<para>
In the Watchmaker Framework, evolutionary operators are defined in terms of the
<interfacename>EvolutionaryOperator</interfacename> interface. This interface
declares a single method that takes a list of selected individuals and returns a
list of evolved individuals. Some operators (i.e. mutation) will process one
individual at a time, whereas others will process individuals in groups
(cross-over processes two individuals at a time).
</para>
<indexterm><primary><classname>StringCrossover</classname></primary></indexterm>
<indexterm><primary><classname>StringMutation</classname></primary></indexterm>
<para>
As with candidate factories, evolutionary operators have associated types that
must be compatible with the type of the evolution engine that they are used with.
And, as with candidate factories, the framework provides several ready-made operators
for common types. These can be found in the
<package>org.uncommons.watchmaker.framework.operators</package> package. The
cross-over and mutation operators that we need for our "Hello World" program are
provided by the <classname>StringCrossover</classname> and
<classname>StringMutation</classname> classes.
</para>
<section>
<title>The Evolution Pipeline</title>
<indexterm significance="preferred"><primary><classname>EvolutionPipeline</classname></primary></indexterm>
<para>
Alert readers will have noticed that the evolution engine constructor only accepts
a single evolutionary operator. So how can we use both cross-over and mutation?
The answer is provided by the <classname>EvolutionPipeline</classname> operator.
This is a compound evolutionary operator that chains together multiple operators of
the same type.
</para>
<informalexample>
<programlisting language="java">
<![CDATA[List<EvolutionaryOperator<String>> operators
= new LinkedList<EvolutionaryOperator<String>>();
operators.add(new StringCrossover());
operators.add(new StringMutation(chars, new Probability(0.02)));
EvolutionaryOperator<String> pipeline
= new EvolutionPipeline<String>(operators);]]>
</programlisting>
</informalexample>
<note>
<para>
The evolution pipeline is just one of many useful operators included
in the <package>org.uncommons.watchmaker.framework.operators</package> package.
Elaborate evolution schemes can be constructed from combinations of these
operators.
Users of the Watchmaker Framework should take a few minutes to browse the API
documentation and familiarise themselves with the available classes.
</para>
</note>
</section>
</section>
<section>
<title>The Fitness Evaluator</title>
<indexterm significance="preferred"><primary><interfacename>FitnessEvaluator</interfacename></primary></indexterm>
<para>
So far we've been able to build our evolutionary program by simply combining instances
of classes provided by the framework. There is one part of the program that we will
always have to write for ourselves though and that is the fitness function, which is
necessarily different for every program.
</para>
<para>
In the Watchmaker Framework, a fitness function is written by implementing the
<interfacename>FitnessEvaluator</interfacename> interface. The
<methodname>getFitness</methodname> method of this interface takes a candidate solution
and returns its fitness score as a Java double. The method actually takes two
arguments, but we can ignore the second for now.
</para>
<para>
The listing below is a fitness evaluator for the "Hello World" program. It
simply assigns one point for each character in the candidate string that
matches the corresponding position in the target string.
</para>
<informalexample>
<programlisting language="java">
<![CDATA[public class StringEvaluator implements FitnessEvaluator<String>
{
private final String targetString = "HELLO WORLD";
/**
* Assigns one "fitness point" for every character in the
* candidate String that matches the corresponding position in
* the target string.
*/
public double getFitness(String candidate,
List<? extends String> population)
{
int matches = 0;
for (int i = 0; i < candidate.length(); i++)
{
if (candidate.charAt(i) == targetString.charAt(i))
{
++matches;
}
}
return matches;
}
public boolean isNatural()
{
return true;
}
}]]>
</programlisting>
</informalexample>
<indexterm><primary>fitness function</primary><secondary>natural</secondary></indexterm>
<indexterm><primary>natural fitness</primary></indexterm>
<para>
By some fitness measures, a higher value indicates a fitter solution. In other
cases a lower value is better. The <methodname>isNatural</methodname> method
of a fitness evaluator simply specifies which scenario applies. In Watchmaker
Framework terminology, a <emphasis>natural</emphasis> fitness function is one that
returns higher values for fitter individuals.
</para>
</section>
<section>
<title>Selection Strategy</title>
<indexterm><primary>selection</primary></indexterm>
<indexterm><primary>SelectionStrategy</primary></indexterm>
<para>
Selection is a key ingredient in any evolutionary algorithm. It's what determines
which individuals survive to reproduce and which are discarded. All we've said about
selection so far is that it should favour fitter individuals. This definition permits
several different implementations. The Watchmaker Framework includes all of the most
common selection strategies in the
<package>org.uncommons.watchmaker.framework.selection</package> package. These are
sufficient for most evolutionary algorithms but, if necessary, it is straightforward
to write your own implementation of the <interfacename>SelectionStrategy</interfacename>
interface.
</para>
<indexterm><primary>RouletteWheelSelection</primary></indexterm>
<para>
Some selection strategies work better than others for certain problems. Often a little
trial-and-error is required to pick the best option. We will delve into the details of
various selection strategies in <xref linkend="selection_chapter" />, but for now we will
just create an instance of the <classname>RouletteWheelSelection</classname> class and use
that for our "Hello World" application.
</para>
<indexterm><primary>fitness-proportionate selection</primary></indexterm>
<indexterm><primary>roulette wheel selection</primary></indexterm>
<indexterm><primary>selection</primary><secondary>fitness-proportionate</secondary></indexterm>
<indexterm><primary>selection</primary><secondary>roulette wheel</secondary></indexterm>
<para>
<emphasis>Roulette wheel selection</emphasis> is the most common type of
<emphasis>fitness-proportionate selection</emphasis>.
It gives all individuals a chance of being selected but favours the fitter
individuals since an individual's selection probability is derived from its
fitness score.
</para>
</section>
<section>
<title>Random Number Generator</title>
<indexterm><primary>random number generator</primary></indexterm>
<indexterm><primary>Random</primary></indexterm>
<indexterm><primary>RNG</primary></indexterm>
<indexterm><primary>SecureRandom</primary></indexterm>
<para>
The final dependency that must be satisfied in order to create an evolution engine
is the random number generator (RNG). An evolution engine has a single RNG that it
passes to its candidate factory, evolutionary operator and selection strategy.
An ideal RNG is both fast and statistically random. We <emphasis>could</emphasis>
use the standard Java RNG, <classname>java.util.Random</classname>, but its output
is not as random as it should be. The other RNG in the standard library,
<classname>java.security.SecureRandom</classname> is much better in this respect
but can be slow.
</para>
<indexterm><primary>MersenneTwisterRNG</primary></indexterm>
<para>
Fortunately, the Watchmaker Framework provides alternatives. The
<classname>org.uncommons.maths.random.MersenneTwisterRNG</classname> random number
generator is both fast and statistically sound. It is usually the best choice
when creating an evolution engine.
</para>
</section>
<section>
<title>Completing the Jigsaw</title>
<para>
We've now got all of the necessary pieces to complete the "Hello World" example
application. Assuming that you've already created the
<classname>StringEvaluator</classname> class (defined above) in a separate file,
the code needed to create the evolution engine looks like this:
</para>
<informalexample>
<programlisting language="java">
<![CDATA[// Create a factory to generate random 11-character Strings.
char[] chars = new char[27];
for (char c = 'A'; c <= 'Z'; c++)
{
chars[c - 'A'] = c;
}
chars[26] = ' ';
CandidateFactory<String> factory = new StringFactory(chars, 11);
// Create a pipeline that applies cross-over then mutation.
List<EvolutionaryOperator<String>> operators
= new LinkedList<EvolutionaryOperator<String>>();
operators.add(new StringCrossover())
operators.add(new StringMutation(chars, new Probability(0.02)));
EvolutionaryOperator<String> pipeline
= new EvolutionPipeline<String>(operators);
FitnessEvaluator<String> fitnessEvaluator = new StringEvaluator();
SelectionStrategy<Object> selection = new RouletteWheelSelection();
Random rng = new MersenneTwisterRNG();
EvolutionEngine<String> engine
= new GenerationalEvolutionEngine<String>(factory,
pipeline,
fitnessEvaluator,
selection,
rng);]]>
</programlisting>
</informalexample>
<indexterm><primary>evolve method</primary></indexterm>
<indexterm><primary>population</primary><secondary>size of</secondary></indexterm>
<para>
The listing above only creates the evolution engine, it does not perform any
evolution. For that we need to call the <methodname>evolve</methodname> method.
The <methodname>evolve</methodname> method takes three parameters. The first
is the size of the population. This is the number of candidate solutions that
exist at any time. A bigger population will often result in a satisfactory
solution being found in fewer generations. On the other hand, the processing
of each generation will take longer because there are more individuals to deal
with. For the "Hello World" program, a population size of 10 is fine.
</para>
<para>
The second parameter is concerned with <emphasis>elitism</emphasis>. Elitism
is explained in <xref linkend="selection_chapter" />. For now, just use a value of zero.
The final varargs parameter specifies one or more termination conditions.
</para>
<section>
<title>Termination Conditions</title>
<indexterm><primary>TerminationCondition</primary></indexterm>
<indexterm><primary>TargetFitness</primary></indexterm>
<para>
Termination conditions make the evolution stop. There are a few reasons why
we would like the evolution to stop. The most obvious is because we have found the
solution that we are looking for. In the case of the "Hello World" program, that
is when we have found the target string. The target string has a fitness score of
11 so we use the <classname>TargetFitness</classname> condition.
</para>
<para>
To complete the evolutionary "Hello World" application, add the following two lines:
</para>
<informalexample>
<programlisting language="java">
<![CDATA[String result = engine.evolve(10, 0, new TargetFitness(11, true));
System.out.println(result);]]>
</programlisting>
</informalexample>
<note>
<indexterm><primary>ElapsedTime</primary></indexterm>
<indexterm><primary>GenerationCount</primary></indexterm>
<indexterm><primary>Stagnation</primary></indexterm>
<para>
When we move on to less trivial evolutionary programs, we will rarely be able to
specify the outcome so precisely. The
<package>org.uncommons.watchmaker.framework.termination</package> package includes
other termination conditions that can be used. For example, we may want the program
to run for a certain period of time, or a certain number of generations, and then
return the best solution it has found up until that point. The
<classname>ElapsedTime</classname> and <classname>GenerationCount</classname>
conditions provide this functionality. Alternatively, we may want the program to
continue as long as it is finding progressively better solutions. The
<classname>Stagnation</classname> condition will terminate the evolution after a
set number of generations pass without any improvement in the fitness of the fittest
candidate.
If multiple termination conditions are specified, the evolution will stop as soon
as any one of them is satisfied.
</para>
</note>
</section>
<section>
<title>Evolution Observers</title>
<para>
Compile and run the above code and, perhaps after a brief pause, you'll see the
following output:
</para>
<informalexample>
<programlisting>
<![CDATA[ HELLO WORLD]]>
</programlisting>
</informalexample>
<indexterm><primary>EvolutionObserver</primary></indexterm>
<para>
This is quite probably the most convoluted "Hello World" program you'll ever write.
It also gives no hints as to its evolutionary nature. We can make the program more
interesting by adding an <interfacename>EvolutionObserver</interfacename> to report
on the progress of the evolution at the end of each generation. Add the following
code to your program before the call to the <methodname>evolve</methodname> method:
</para>
<informalexample>
<programlisting language="java">
<![CDATA[engine.addEvolutionObserver(new EvolutionObserver<String>()
{
public void populationUpdate(PopulationData<? extends String> data)
{
System.out.printf("Generation %d: %s\n",
data.getGenerationNumber(),
data.getBestCandidate());
}
});]]>
</programlisting>
</informalexample>
<para>
Re-compile the program and run it again. This time you'll see all of the steps
taken to arrive at the target string:
</para>
<informalexample>
<programlisting>
Generation 0: JIKDORHOQZJ
Generation 1: ULLLFQWZPXG
Generation 2: UEULKFVFZLS
Generation 3: KLLLFKZGRLS
Generation 4: HLLLFKZGRLS
Generation 5: HEDPOYWOZLS
Generation 6: HEULKIWWZLD
Generation 7: HPRLOYWOZLS
Generation 8: HEULOYWOZLS
Generation 9: HEULOYWORLS
Generation 10: HEULOYWORLS
Generation 11: HPLLK WQRLH
Generation 12: HEBLOYWQRLS
Generation 13: HEULOYWOBLA
Generation 14: HEBLOIWMRLD
Generation 15: HEBLOIWMRLD
Generation 16: HEYLFNWQRLD
Generation 17: HEBLOIWORLS
Generation 18: HEBLOIWORLT
Generation 19: HEBLOKWGRLD
Generation 20: HELLAYWORLS
Generation 21: HELHOIWORLT
Generation 22: HEWLOIWORLS
Generation 23: HEBLOYCORLD
Generation 24: HELLKQWORLD
Generation 25: HELLOIWORLT
Generation 26: HELLOIWORLS
Generation 27: HELLKQWORLD
Generation 28: HELLFYWORLD
Generation 29: HELLOIWORLD
Generation 30: HELLOIWORLD
Generation 31: HELLOIWORLD
Generation 32: HELLOIWORLD
Generation 33: HELLOIWORLD
Generation 34: HELLOIWORLD
Generation 35: HELLOIWDRLD
Generation 36: HELLOIWORLD
Generation 37: HELLOIWORLD
Generation 38: HELLOPWORLD
Generation 39: HELLOIWORLD
Generation 40: HELLO WORLD
HELLO WORLD
</programlisting>
</informalexample>
</section>
</section>
</chapter>
|