[11770] | 1 | using System;
|
---|
| 2 | using System.Collections.Generic;
|
---|
| 3 | using System.Diagnostics;
|
---|
| 4 | using System.Linq;
|
---|
| 5 | using System.Resources;
|
---|
| 6 | using System.Runtime.InteropServices;
|
---|
| 7 | using System.Text;
|
---|
| 8 | using HeuristicLab.Algorithms.Bandits;
|
---|
| 9 | using HeuristicLab.Algorithms.Bandits.BanditPolicies;
|
---|
| 10 | using HeuristicLab.Algorithms.Bandits.GrammarPolicies;
|
---|
| 11 | using HeuristicLab.Common;
|
---|
| 12 | using HeuristicLab.Problems.GrammaticalOptimization;
|
---|
| 13 |
|
---|
| 14 | namespace HeuristicLab.Algorithms.GrammaticalOptimization {
|
---|
| 15 | // a search procedure that uses a policy to generate sentences and updates the policy (online RL)
|
---|
| 16 | // 1) Start with phrase = sentence symbol of grammar
|
---|
| 17 | // 2) Repeat
|
---|
| 18 | // a) generate derived phrases using left-canonical derivation and grammar rules
|
---|
| 19 | // b) keep only the phrases which are allowed (sentence length limit)
|
---|
| 20 | // c) if the set of phrases is empty restart with 1)
|
---|
| 21 | // d) otherwise use policy to select one of the possible derived phrases as active phrase
|
---|
| 22 | // the policy has the option to fail (for instance if all derived phrases are terminal and should not be visited again), in this case we restart at 1
|
---|
| 23 | // ... until phrase is terminal
|
---|
| 24 | // 3) Collect reward and update policy (feedback: state of visited rewards from step 2)
|
---|
[11846] | 25 | public class SequentialSearch : SolverBase {
|
---|
[11793] | 26 | // only for storing states so that it is not necessary to allocate new state strings whenever we select a follow state using the policy
|
---|
| 27 | private class TreeNode {
|
---|
| 28 | public int randomTries;
|
---|
| 29 | public string phrase;
|
---|
| 30 | public Sequence alternative;
|
---|
| 31 | public TreeNode[] children;
|
---|
[11770] | 32 |
|
---|
[11793] | 33 | public TreeNode(string phrase, Sequence alternative) {
|
---|
| 34 | this.alternative = alternative;
|
---|
| 35 | this.phrase = phrase;
|
---|
| 36 | }
|
---|
| 37 | }
|
---|
| 38 |
|
---|
| 39 |
|
---|
[11770] | 40 | private readonly int maxLen;
|
---|
| 41 | private readonly IProblem problem;
|
---|
| 42 | private readonly Random random;
|
---|
| 43 | private readonly int randomTries;
|
---|
| 44 | private readonly IGrammarPolicy behaviourPolicy;
|
---|
| 45 | private readonly IGrammarPolicy greedyPolicy;
|
---|
[11793] | 46 | private TreeNode rootNode;
|
---|
| 47 |
|
---|
| 48 | private int tries;
|
---|
[11770] | 49 | private int maxSearchDepth;
|
---|
| 50 |
|
---|
| 51 | private string bestPhrase;
|
---|
[11793] | 52 | private readonly List<string> stateChain;
|
---|
[11770] | 53 |
|
---|
| 54 | public SequentialSearch(IProblem problem, int maxLen, Random random, int randomTries, IGrammarPolicy behaviourPolicy) {
|
---|
| 55 | this.maxLen = maxLen;
|
---|
| 56 | this.problem = problem;
|
---|
| 57 | this.random = random;
|
---|
| 58 | this.randomTries = randomTries;
|
---|
| 59 | this.behaviourPolicy = behaviourPolicy;
|
---|
[11793] | 60 | this.greedyPolicy = new GenericGrammarPolicy(problem, new EpsGreedyPolicy(0.0), false);
|
---|
| 61 | this.stateChain = new List<string>();
|
---|
[11770] | 62 | }
|
---|
| 63 |
|
---|
[11977] | 64 | public bool StopRequested {
|
---|
| 65 | get;
|
---|
| 66 | set;
|
---|
| 67 | }
|
---|
| 68 |
|
---|
[11846] | 69 | public override void Run(int maxIterations) {
|
---|
[11770] | 70 | Reset();
|
---|
| 71 |
|
---|
[11977] | 72 | for (int i = 0; !StopRequested && !Done() && i < maxIterations; i++) {
|
---|
[11770] | 73 | var phrase = SampleSentence(problem.Grammar);
|
---|
| 74 | // can fail on the last sentence
|
---|
| 75 | if (phrase.IsTerminal) {
|
---|
| 76 | var sentence = phrase.ToString();
|
---|
| 77 | tries++;
|
---|
| 78 | var quality = problem.Evaluate(sentence) / problem.BestKnownQuality(maxLen);
|
---|
[11806] | 79 | if (double.IsNaN(quality)) quality = 0.0;
|
---|
[11770] | 80 | Debug.Assert(quality >= 0 && quality <= 1.0);
|
---|
| 81 |
|
---|
| 82 | if (quality > bestQuality) {
|
---|
| 83 | bestPhrase = sentence;
|
---|
| 84 | }
|
---|
[11846] | 85 |
|
---|
| 86 | OnSolutionEvaluated(sentence, quality);
|
---|
| 87 | DistributeReward(quality);
|
---|
| 88 |
|
---|
[11770] | 89 | }
|
---|
| 90 | }
|
---|
| 91 | }
|
---|
| 92 |
|
---|
| 93 |
|
---|
[11793] | 94 | private Sequence SampleSentence(IGrammar grammar) {
|
---|
| 95 | Sequence phrase;
|
---|
[11770] | 96 | do {
|
---|
| 97 | stateChain.Clear();
|
---|
[11793] | 98 | phrase = new Sequence(rootNode.phrase);
|
---|
[11770] | 99 | } while (!Done() && !TryCompleteSentence(grammar, ref phrase));
|
---|
| 100 | return phrase;
|
---|
| 101 | }
|
---|
| 102 |
|
---|
[11793] | 103 | private bool TryCompleteSentence(IGrammar g, ref Sequence phrase) {
|
---|
[11770] | 104 | if (phrase.Length > maxLen) throw new ArgumentException();
|
---|
| 105 | if (g.MinPhraseLength(phrase) > maxLen) throw new ArgumentException();
|
---|
| 106 | var curDepth = 0;
|
---|
[11793] | 107 | var n = rootNode;
|
---|
| 108 | stateChain.Add(n.phrase);
|
---|
[11770] | 109 |
|
---|
| 110 | while (!phrase.IsTerminal) {
|
---|
[11799] | 111 | if (n.randomTries < randomTries) {
|
---|
| 112 | n.randomTries++;
|
---|
| 113 | maxSearchDepth = Math.Max(maxSearchDepth, curDepth);
|
---|
| 114 | g.CompleteSentenceRandomly(random, phrase, maxLen);
|
---|
| 115 | return true;
|
---|
| 116 | } else {
|
---|
| 117 | // => select using bandit policy
|
---|
| 118 | // failure means we simply restart
|
---|
| 119 | GenerateFollowStates(n); // creates child nodes for node n
|
---|
[11770] | 120 |
|
---|
[11799] | 121 | int selectedChildIdx;
|
---|
| 122 | if (!behaviourPolicy.TrySelect(random, n.phrase, n.children.Select(ch => ch.phrase), out selectedChildIdx)) {
|
---|
| 123 | return false;
|
---|
| 124 | }
|
---|
| 125 | phrase.ReplaceAt(phrase.FirstNonTerminalIndex, 1, n.children[selectedChildIdx].alternative);
|
---|
[11793] | 126 |
|
---|
[11799] | 127 | // prepare for next iteration
|
---|
| 128 | n = n.children[selectedChildIdx];
|
---|
| 129 | stateChain.Add(n.phrase);
|
---|
| 130 | curDepth++;
|
---|
[11793] | 131 | }
|
---|
[11770] | 132 | } // while
|
---|
| 133 |
|
---|
| 134 | maxSearchDepth = Math.Max(maxSearchDepth, curDepth);
|
---|
| 135 | return true;
|
---|
| 136 | }
|
---|
| 137 |
|
---|
| 138 |
|
---|
[11793] | 139 | private IEnumerable<string> GenerateFollowStates(TreeNode n) {
|
---|
| 140 | // create children on the first visit
|
---|
| 141 | if (n.children == null) {
|
---|
| 142 | var g = problem.Grammar;
|
---|
| 143 | // tree is only used for easily retrieving the follow-states of a state
|
---|
| 144 | var phrase = new Sequence(n.phrase);
|
---|
[11770] | 145 | char nt = phrase.FirstNonTerminal;
|
---|
| 146 |
|
---|
| 147 | int maxLenOfReplacement = maxLen - (phrase.Length - 1);
|
---|
| 148 | // replacing aAb with maxLen 4 means we can only use alternatives with a minPhraseLen <= 2
|
---|
| 149 | Debug.Assert(maxLenOfReplacement > 0);
|
---|
| 150 |
|
---|
| 151 | var alts = g.GetAlternatives(nt).Where(alt => g.MinPhraseLength(alt) <= maxLenOfReplacement);
|
---|
| 152 |
|
---|
[11793] | 153 | var children = new TreeNode[alts.Count()];
|
---|
[11770] | 154 | int idx = 0;
|
---|
| 155 | foreach (var alt in alts) {
|
---|
[11799] | 156 | // var newPhrase = new Sequence(phrase); // clone
|
---|
| 157 | // newPhrase.ReplaceAt(newPhrase.FirstNonTerminalIndex, 1, alt);
|
---|
| 158 | // children[idx++] = new TreeNode(newPhrase.ToString(), alt);
|
---|
| 159 |
|
---|
| 160 | // since we are not using a sequence later on we might directly transform the current sequence to a string and replace there
|
---|
| 161 | var phraseStr = phrase.ToString();
|
---|
| 162 | var sb = new StringBuilder(phraseStr);
|
---|
| 163 | sb.Remove(phrase.FirstNonTerminalIndex, 1).Insert(phrase.FirstNonTerminalIndex, alt.ToString());
|
---|
| 164 | children[idx++] = new TreeNode(sb.ToString(), alt);
|
---|
[11770] | 165 | }
|
---|
[11793] | 166 | n.children = children;
|
---|
| 167 | }
|
---|
| 168 | return n.children.Select(ch => ch.phrase);
|
---|
[11770] | 169 | }
|
---|
| 170 |
|
---|
| 171 | private void DistributeReward(double reward) {
|
---|
| 172 | behaviourPolicy.UpdateReward(stateChain, reward);
|
---|
[11976] | 173 | //greedyPolicy.UpdateReward(stateChain, reward);
|
---|
[11770] | 174 | }
|
---|
| 175 |
|
---|
| 176 |
|
---|
| 177 | private void Reset() {
|
---|
[11977] | 178 | StopRequested = false;
|
---|
[11770] | 179 | behaviourPolicy.Reset();
|
---|
| 180 | greedyPolicy.Reset();
|
---|
| 181 | maxSearchDepth = 0;
|
---|
| 182 | bestQuality = 0.0;
|
---|
| 183 | tries = 0;
|
---|
[11793] | 184 | rootNode = new TreeNode(problem.Grammar.SentenceSymbol.ToString(), new ReadonlySequence("$"));
|
---|
[11770] | 185 | }
|
---|
| 186 |
|
---|
| 187 | public bool Done() {
|
---|
[11793] | 188 | int selectedStateIdx;
|
---|
| 189 | return !behaviourPolicy.TrySelect(random, rootNode.phrase, GenerateFollowStates(rootNode), out selectedStateIdx);
|
---|
[11770] | 190 | }
|
---|
| 191 |
|
---|
| 192 | #region introspection
|
---|
| 193 | public void PrintStats() {
|
---|
| 194 | Console.WriteLine("depth: {0,5} tries: {1,5} best phrase {2,50} bestQ {3:F3}", maxSearchDepth, tries, bestPhrase, bestQuality);
|
---|
| 195 |
|
---|
[11793] | 196 | // use behaviour strategy to generate the currently prefered sentence
|
---|
[11770] | 197 | var policy = behaviourPolicy;
|
---|
[11793] | 198 |
|
---|
| 199 | var n = rootNode;
|
---|
| 200 |
|
---|
| 201 | while (n != null) {
|
---|
| 202 | var phrase = n.phrase;
|
---|
[11770] | 203 | Console.ForegroundColor = ConsoleColor.White;
|
---|
| 204 | Console.WriteLine("{0,-30}", phrase);
|
---|
[11793] | 205 | var children = n.children;
|
---|
| 206 | if (children == null || !children.Any()) break;
|
---|
| 207 | var values = children.Select(ch => policy.GetValue(ch.phrase));
|
---|
[11770] | 208 | var maxValue = values.Max();
|
---|
| 209 | if (maxValue == 0) maxValue = 1.0;
|
---|
| 210 |
|
---|
| 211 | // write phrases
|
---|
[11793] | 212 | foreach (var ch in children) {
|
---|
| 213 | SetColorForValue(policy.GetValue(ch.phrase) / maxValue);
|
---|
| 214 | Console.Write(" {0,-4}", ch.phrase.Substring(Math.Max(0, ch.phrase.Length - 3), Math.Min(3, ch.phrase.Length)));
|
---|
[11770] | 215 | }
|
---|
| 216 | Console.WriteLine();
|
---|
| 217 |
|
---|
| 218 | // write values
|
---|
[11793] | 219 | foreach (var ch in children) {
|
---|
| 220 | SetColorForValue(policy.GetValue(ch.phrase) / maxValue);
|
---|
| 221 | Console.Write(" {0:F2}", policy.GetValue(ch.phrase) * 10.0);
|
---|
[11770] | 222 | }
|
---|
| 223 | Console.WriteLine();
|
---|
| 224 |
|
---|
| 225 | // write tries
|
---|
[11793] | 226 | foreach (var ch in children) {
|
---|
| 227 | SetColorForValue(policy.GetValue(ch.phrase) / maxValue);
|
---|
| 228 | Console.Write(" {0,4}", policy.GetTries(ch.phrase));
|
---|
[11770] | 229 | }
|
---|
| 230 | Console.WriteLine();
|
---|
[11793] | 231 | int selectedChildIdx;
|
---|
| 232 | if (!policy.TrySelect(random, phrase, children.Select(ch => ch.phrase), out selectedChildIdx)) {
|
---|
[11770] | 233 | break;
|
---|
| 234 | }
|
---|
[11793] | 235 | n = n.children[selectedChildIdx];
|
---|
[11770] | 236 | }
|
---|
| 237 |
|
---|
| 238 | Console.ForegroundColor = ConsoleColor.White;
|
---|
| 239 | Console.WriteLine("-------------------");
|
---|
| 240 | }
|
---|
| 241 |
|
---|
| 242 | private void SetColorForValue(double v) {
|
---|
| 243 | Console.ForegroundColor = ConsoleEx.ColorForValue(v);
|
---|
| 244 | }
|
---|
| 245 | #endregion
|
---|
[11977] | 246 |
|
---|
[11770] | 247 | }
|
---|
| 248 | }
|
---|