1 | using System;
|
---|
2 | using System.Collections.Generic;
|
---|
3 | using System.Data;
|
---|
4 | using System.Diagnostics;
|
---|
5 | using System.Globalization;
|
---|
6 | using System.Linq;
|
---|
7 | using System.Text;
|
---|
8 | using System.Threading.Tasks;
|
---|
9 | using HeuristicLab.Algorithms.Bandits;
|
---|
10 | using HeuristicLab.Algorithms.Bandits.Models;
|
---|
11 | using HeuristicLab.Algorithms.GrammaticalOptimization;
|
---|
12 | using HeuristicLab.Problems.GrammaticalOptimization;
|
---|
13 |
|
---|
14 | namespace Main {
|
---|
15 | class Program {
|
---|
16 | static void Main(string[] args) {
|
---|
17 | CultureInfo.DefaultThreadCurrentCulture = CultureInfo.InvariantCulture;
|
---|
18 |
|
---|
19 | RunDemo();
|
---|
20 | //RunGridTest();
|
---|
21 | }
|
---|
22 |
|
---|
23 | private static void RunGridTest() {
|
---|
24 | int maxIterations = 100000; // for poly-10 with 50000 evaluations no successful try with hl yet
|
---|
25 | // var globalRandom = new Random(31415);
|
---|
26 | var localRandSeed = 31415;
|
---|
27 | var reps = 20;
|
---|
28 |
|
---|
29 | var policyFactories = new Func<Random, int, IPolicy>[]
|
---|
30 | {
|
---|
31 | (rand, numActions) => new GaussianThompsonSamplingPolicy(rand, numActions),
|
---|
32 | (rand, numActions) => new BernoulliThompsonSamplingPolicy(rand, numActions),
|
---|
33 | (rand, numActions) => new RandomPolicy(rand, numActions),
|
---|
34 | (rand, numActions) => new EpsGreedyPolicy(rand, numActions, 0.01),
|
---|
35 | (rand, numActions) => new EpsGreedyPolicy(rand, numActions, 0.05),
|
---|
36 | (rand, numActions) => new EpsGreedyPolicy(rand, numActions, 0.1),
|
---|
37 | (rand, numActions) => new EpsGreedyPolicy(rand, numActions, 0.2),
|
---|
38 | (rand, numActions) => new EpsGreedyPolicy(rand, numActions, 0.5),
|
---|
39 | (rand, numActions) => new UCTPolicy(numActions, 0.1),
|
---|
40 | (rand, numActions) => new UCTPolicy(numActions, 0.5),
|
---|
41 | (rand, numActions) => new UCTPolicy(numActions, 1),
|
---|
42 | (rand, numActions) => new UCTPolicy(numActions, 2),
|
---|
43 | (rand, numActions) => new UCTPolicy(numActions, 5),
|
---|
44 | (rand, numActions) => new UCTPolicy(numActions, 10),
|
---|
45 | (rand, numActions) => new UCB1Policy(numActions),
|
---|
46 | (rand, numActions) => new UCB1TunedPolicy(numActions),
|
---|
47 | (rand, numActions) => new UCBNormalPolicy(numActions),
|
---|
48 | (rand, numActions) => new BoltzmannExplorationPolicy(rand, numActions, 0.1),
|
---|
49 | (rand, numActions) => new BoltzmannExplorationPolicy(rand, numActions, 0.5),
|
---|
50 | (rand, numActions) => new BoltzmannExplorationPolicy(rand, numActions, 1),
|
---|
51 | (rand, numActions) => new BoltzmannExplorationPolicy(rand, numActions, 5),
|
---|
52 | (rand, numActions) => new BoltzmannExplorationPolicy(rand, numActions, 10),
|
---|
53 | (rand, numActions) => new BoltzmannExplorationPolicy(rand, numActions, 20),
|
---|
54 | (rand, numActions) => new BoltzmannExplorationPolicy(rand, numActions, 100),
|
---|
55 | (rand, numActions) => new ChernoffIntervalEstimationPolicy(numActions, 0.01),
|
---|
56 | (rand, numActions) => new ChernoffIntervalEstimationPolicy(numActions, 0.05),
|
---|
57 | (rand, numActions) => new ChernoffIntervalEstimationPolicy(numActions, 0.1),
|
---|
58 | (rand, numActions) => new ChernoffIntervalEstimationPolicy(numActions, 0.2),
|
---|
59 | (rand, numActions) => new ThresholdAscentPolicy(numActions, 10, 0.01),
|
---|
60 | (rand, numActions) => new ThresholdAscentPolicy(numActions, 10, 0.05),
|
---|
61 | (rand, numActions) => new ThresholdAscentPolicy(numActions, 10, 0.1),
|
---|
62 | (rand, numActions) => new ThresholdAscentPolicy(numActions, 10, 0.2),
|
---|
63 | (rand, numActions) => new ThresholdAscentPolicy(numActions, 100, 0.01),
|
---|
64 | (rand, numActions) => new ThresholdAscentPolicy(numActions, 100, 0.05),
|
---|
65 | (rand, numActions) => new ThresholdAscentPolicy(numActions, 100, 0.1),
|
---|
66 | (rand, numActions) => new ThresholdAscentPolicy(numActions, 100, 0.2),
|
---|
67 | (rand, numActions) => new ThresholdAscentPolicy(numActions, 1000, 0.01),
|
---|
68 | (rand, numActions) => new ThresholdAscentPolicy(numActions, 1000, 0.05),
|
---|
69 | (rand, numActions) => new ThresholdAscentPolicy(numActions, 1000, 0.1),
|
---|
70 | (rand, numActions) => new ThresholdAscentPolicy(numActions, 1000, 0.2),
|
---|
71 | (rand, numActions) => new ThresholdAscentPolicy(numActions, 5000, 0.01),
|
---|
72 | (rand, numActions) => new ThresholdAscentPolicy(numActions, 10000, 0.01),
|
---|
73 | };
|
---|
74 |
|
---|
75 | var tasks = new List<Task>();
|
---|
76 | foreach (var randomTries in new int[] { 1, 10, /* 5, 100 /*, 500, 1000 */}) {
|
---|
77 | foreach (var policyFactory in policyFactories) {
|
---|
78 | var myPolicyFactory = policyFactory;
|
---|
79 | var myRandomTries = randomTries;
|
---|
80 | var localRand = new Random(localRandSeed);
|
---|
81 | var options = new ParallelOptions();
|
---|
82 | options.MaxDegreeOfParallelism = 1;
|
---|
83 | Parallel.For(0, reps, options, (i) => {
|
---|
84 | //var t = Task.Run(() => {
|
---|
85 | Random myLocalRand;
|
---|
86 | lock (localRand)
|
---|
87 | myLocalRand = new Random(localRand.Next());
|
---|
88 |
|
---|
89 | //for (int i = 0; i < reps; i++) {
|
---|
90 |
|
---|
91 | int iterations = 0;
|
---|
92 | var sw = new Stopwatch();
|
---|
93 | var globalStatistics = new SentenceSetStatistics();
|
---|
94 |
|
---|
95 | var problem = new SymbolicRegressionPoly10Problem();
|
---|
96 | //var problem = new SantaFeAntProblem();
|
---|
97 | //var problem = new PalindromeProblem();
|
---|
98 | //var problem = new HardPalindromeProblem();
|
---|
99 | //var problem = new RoyalPairProblem();
|
---|
100 | //var problem = new EvenParityProblem();
|
---|
101 | var alg = new MctsSampler(problem, 25, myLocalRand, myRandomTries, myPolicyFactory);
|
---|
102 | //var alg = new ExhaustiveBreadthFirstSearch(problem, 25);
|
---|
103 | //var alg = new AlternativesContextSampler(problem, 25);
|
---|
104 |
|
---|
105 | alg.SolutionEvaluated += (sentence, quality) => {
|
---|
106 | iterations++;
|
---|
107 | globalStatistics.AddSentence(sentence, quality);
|
---|
108 | if (iterations % 10000 == 0) {
|
---|
109 | Console.WriteLine("{0,4} {1,7} {2,5} {3,25} {4}", alg.treeDepth, alg.treeSize, myRandomTries, myPolicyFactory(myLocalRand, 1), globalStatistics);
|
---|
110 | }
|
---|
111 | };
|
---|
112 |
|
---|
113 | sw.Start();
|
---|
114 |
|
---|
115 | alg.Run(maxIterations);
|
---|
116 |
|
---|
117 | sw.Stop();
|
---|
118 | //Console.WriteLine("{0,5} {1} {2}", randomTries, policyFactory(1), globalStatistics);
|
---|
119 | //}
|
---|
120 | //});
|
---|
121 | //tasks.Add(t);
|
---|
122 | });
|
---|
123 | }
|
---|
124 | }
|
---|
125 | //Task.WaitAll(tasks.ToArray());
|
---|
126 | }
|
---|
127 |
|
---|
128 | private static void RunDemo() {
|
---|
129 | // TODO: warum funktioniert die alte Implementierung von GaussianThompson besser für SantaFe als alte? Siehe Vergleich: alte vs. neue implementierung GaussianThompsonSampling
|
---|
130 | // TODO: why does GaussianThompsonSampling work so well with MCTS for the artificial ant problem?
|
---|
131 | // TODO: wie kann ich sampler noch vergleichen bzw. was kann man messen um die qualität des samplers abzuschätzen (bis auf qualität und iterationen bis zur besten lösung) => ziel schnellere iterationen zu gutem ergebnis
|
---|
132 | // TODO: likelihood für R=1 bei Gaussian oder GaussianMixture einfach berechenbar?
|
---|
133 | // TODO: research thompson sampling for max bandit?
|
---|
134 | // TODO: ausführlicher test von strategien für k-armed max bandit
|
---|
135 | // TODO: verify TA implementation using example from the original paper
|
---|
136 | // TODO: reference HL.ProblemInstances and try on tower dataset
|
---|
137 | // TODO: compare results for different policies also for the symb-reg problem
|
---|
138 | // TODO: separate policy from MCTS tree data structure to allow sharing of information over disconnected parts of the tree (semantic equivalence)
|
---|
139 | // TODO: implement thompson sampling for gaussian mixture models
|
---|
140 | // TODO: implement inspection for MCTS (eventuell interactive command line für statistiken aus dem baum anzeigen)
|
---|
141 | // TODO: implement ACO-style bandit policy
|
---|
142 | // TODO: implement sequences that can be manipulated in-place (instead of strings), alternatives are also stored as sequences, for a sequence the index of the first NT-symb can be stored
|
---|
143 | // TODO: gleichzeitige modellierung von transformierter zielvariable (y, 1/y, log(y), exp(y), sqrt(y), ...)
|
---|
144 | // TODO: vergleich bei complete-randomly möglichst kurze sätze generieren vs. einfach zufällig alternativen wählen
|
---|
145 | // TODO: reward discounting (für veränderliche reward distributions über zeit). speziellen unit-test dafür erstellen
|
---|
146 |
|
---|
147 |
|
---|
148 | int maxIterations = 10000000;
|
---|
149 | int iterations = 0;
|
---|
150 | var sw = new Stopwatch();
|
---|
151 | double bestQuality = 0;
|
---|
152 | string bestSentence = "";
|
---|
153 | var globalStatistics = new SentenceSetStatistics();
|
---|
154 | var random = new Random();
|
---|
155 |
|
---|
156 | //var problem = new SymbolicRegressionPoly10Problem();
|
---|
157 | var problem = new SantaFeAntProblem();
|
---|
158 | //var problem = new PalindromeProblem();
|
---|
159 | //var problem = new HardPalindromeProblem();
|
---|
160 | //var problem = new RoyalPairProblem();
|
---|
161 | //var problem = new EvenParityProblem();
|
---|
162 | //var alg = new MctsSampler(problem, 17, random, 1, (rand, numActions) => new GenericThompsonSamplingPolicy(rand, numActions, new GaussianModel(numActions, 0.5, 10)));
|
---|
163 | //var alg = new ExhaustiveBreadthFirstSearch(problem, 17);
|
---|
164 | //var alg = new AlternativesContextSampler(problem, random, 17, 4, (rand, numActions) => new RandomPolicy(rand, numActions));
|
---|
165 | //var alg = new ExhaustiveDepthFirstSearch(problem, 17);
|
---|
166 | // var alg = new AlternativesSampler(problem, 17);
|
---|
167 | var alg = new RandomSearch(problem, random, 17);
|
---|
168 |
|
---|
169 | alg.FoundNewBestSolution += (sentence, quality) => {
|
---|
170 | bestQuality = quality;
|
---|
171 | bestSentence = sentence;
|
---|
172 | Console.WriteLine("{0,10} {1,10:F5} {2,10:F5} {3}", iterations, bestQuality, quality, sentence);
|
---|
173 | };
|
---|
174 | alg.SolutionEvaluated += (sentence, quality) => {
|
---|
175 | iterations++;
|
---|
176 | globalStatistics.AddSentence(sentence, quality);
|
---|
177 | if (iterations % 1000 == 0) {
|
---|
178 | //alg.PrintStats();
|
---|
179 | }
|
---|
180 | if (iterations % 10000 == 0) {
|
---|
181 | //Console.WriteLine("{0,10} {1,10:F5} {2,10:F5} {3}", iterations, bestQuality, quality, sentence);
|
---|
182 | //Console.WriteLine("{0,4} {1,7} {2}", alg.treeDepth, alg.treeSize, globalStatistics);
|
---|
183 | Console.WriteLine(globalStatistics);
|
---|
184 | }
|
---|
185 | };
|
---|
186 |
|
---|
187 |
|
---|
188 | sw.Start();
|
---|
189 |
|
---|
190 | alg.Run(maxIterations);
|
---|
191 |
|
---|
192 | sw.Stop();
|
---|
193 |
|
---|
194 | Console.WriteLine("{0,10} Best soultion: {1,10:F5} {2}", iterations, bestQuality, bestSentence);
|
---|
195 | Console.WriteLine("{0:F2} sec {1,10:F1} sols/sec {2,10:F1} ns/sol",
|
---|
196 | sw.Elapsed.TotalSeconds,
|
---|
197 | maxIterations / (double)sw.Elapsed.TotalSeconds,
|
---|
198 | (double)sw.ElapsedMilliseconds * 1000 / maxIterations);
|
---|
199 | }
|
---|
200 | }
|
---|
201 | }
|
---|