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)
|
---|
25 | public class SequentialSearch : SolverBase {
|
---|
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;
|
---|
32 |
|
---|
33 | public TreeNode(string phrase, Sequence alternative) {
|
---|
34 | this.alternative = alternative;
|
---|
35 | this.phrase = phrase;
|
---|
36 | }
|
---|
37 | }
|
---|
38 |
|
---|
39 |
|
---|
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;
|
---|
46 | private TreeNode rootNode;
|
---|
47 |
|
---|
48 | private int tries;
|
---|
49 | private int maxSearchDepth;
|
---|
50 |
|
---|
51 | private string bestPhrase;
|
---|
52 | private readonly List<string> stateChain;
|
---|
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;
|
---|
60 | this.greedyPolicy = new GenericGrammarPolicy(problem, new EpsGreedyPolicy(0.0), false);
|
---|
61 | this.stateChain = new List<string>();
|
---|
62 | }
|
---|
63 |
|
---|
64 | public override void Run(int maxIterations) {
|
---|
65 | Reset();
|
---|
66 |
|
---|
67 | for (int i = 0; /*!bestQuality.IsAlmost(1.0) && */!Done() && i < maxIterations; i++) {
|
---|
68 | var phrase = SampleSentence(problem.Grammar);
|
---|
69 | // can fail on the last sentence
|
---|
70 | if (phrase.IsTerminal) {
|
---|
71 | var sentence = phrase.ToString();
|
---|
72 | tries++;
|
---|
73 | var quality = problem.Evaluate(sentence) / problem.BestKnownQuality(maxLen);
|
---|
74 | if (double.IsNaN(quality)) quality = 0.0;
|
---|
75 | Debug.Assert(quality >= 0 && quality <= 1.0);
|
---|
76 |
|
---|
77 | if (quality > bestQuality) {
|
---|
78 | bestPhrase = sentence;
|
---|
79 | }
|
---|
80 |
|
---|
81 | OnSolutionEvaluated(sentence, quality);
|
---|
82 | DistributeReward(quality);
|
---|
83 |
|
---|
84 | }
|
---|
85 | }
|
---|
86 | }
|
---|
87 |
|
---|
88 |
|
---|
89 | private Sequence SampleSentence(IGrammar grammar) {
|
---|
90 | Sequence phrase;
|
---|
91 | do {
|
---|
92 | stateChain.Clear();
|
---|
93 | phrase = new Sequence(rootNode.phrase);
|
---|
94 | } while (!Done() && !TryCompleteSentence(grammar, ref phrase));
|
---|
95 | return phrase;
|
---|
96 | }
|
---|
97 |
|
---|
98 | private bool TryCompleteSentence(IGrammar g, ref Sequence phrase) {
|
---|
99 | if (phrase.Length > maxLen) throw new ArgumentException();
|
---|
100 | if (g.MinPhraseLength(phrase) > maxLen) throw new ArgumentException();
|
---|
101 | var curDepth = 0;
|
---|
102 | var n = rootNode;
|
---|
103 | stateChain.Add(n.phrase);
|
---|
104 |
|
---|
105 | while (!phrase.IsTerminal) {
|
---|
106 | if (n.randomTries < randomTries) {
|
---|
107 | n.randomTries++;
|
---|
108 | maxSearchDepth = Math.Max(maxSearchDepth, curDepth);
|
---|
109 | g.CompleteSentenceRandomly(random, phrase, maxLen);
|
---|
110 | return true;
|
---|
111 | } else {
|
---|
112 | // => select using bandit policy
|
---|
113 | // failure means we simply restart
|
---|
114 | GenerateFollowStates(n); // creates child nodes for node n
|
---|
115 |
|
---|
116 | int selectedChildIdx;
|
---|
117 | if (!behaviourPolicy.TrySelect(random, n.phrase, n.children.Select(ch => ch.phrase), out selectedChildIdx)) {
|
---|
118 | return false;
|
---|
119 | }
|
---|
120 | phrase.ReplaceAt(phrase.FirstNonTerminalIndex, 1, n.children[selectedChildIdx].alternative);
|
---|
121 |
|
---|
122 | // prepare for next iteration
|
---|
123 | n = n.children[selectedChildIdx];
|
---|
124 | stateChain.Add(n.phrase);
|
---|
125 | curDepth++;
|
---|
126 | }
|
---|
127 | } // while
|
---|
128 |
|
---|
129 | maxSearchDepth = Math.Max(maxSearchDepth, curDepth);
|
---|
130 | return true;
|
---|
131 | }
|
---|
132 |
|
---|
133 |
|
---|
134 | private IEnumerable<string> GenerateFollowStates(TreeNode n) {
|
---|
135 | // create children on the first visit
|
---|
136 | if (n.children == null) {
|
---|
137 | var g = problem.Grammar;
|
---|
138 | // tree is only used for easily retrieving the follow-states of a state
|
---|
139 | var phrase = new Sequence(n.phrase);
|
---|
140 | char nt = phrase.FirstNonTerminal;
|
---|
141 |
|
---|
142 | int maxLenOfReplacement = maxLen - (phrase.Length - 1);
|
---|
143 | // replacing aAb with maxLen 4 means we can only use alternatives with a minPhraseLen <= 2
|
---|
144 | Debug.Assert(maxLenOfReplacement > 0);
|
---|
145 |
|
---|
146 | var alts = g.GetAlternatives(nt).Where(alt => g.MinPhraseLength(alt) <= maxLenOfReplacement);
|
---|
147 |
|
---|
148 | var children = new TreeNode[alts.Count()];
|
---|
149 | int idx = 0;
|
---|
150 | foreach (var alt in alts) {
|
---|
151 | // var newPhrase = new Sequence(phrase); // clone
|
---|
152 | // newPhrase.ReplaceAt(newPhrase.FirstNonTerminalIndex, 1, alt);
|
---|
153 | // children[idx++] = new TreeNode(newPhrase.ToString(), alt);
|
---|
154 |
|
---|
155 | // since we are not using a sequence later on we might directly transform the current sequence to a string and replace there
|
---|
156 | var phraseStr = phrase.ToString();
|
---|
157 | var sb = new StringBuilder(phraseStr);
|
---|
158 | sb.Remove(phrase.FirstNonTerminalIndex, 1).Insert(phrase.FirstNonTerminalIndex, alt.ToString());
|
---|
159 | children[idx++] = new TreeNode(sb.ToString(), alt);
|
---|
160 | }
|
---|
161 | n.children = children;
|
---|
162 | }
|
---|
163 | return n.children.Select(ch => ch.phrase);
|
---|
164 | }
|
---|
165 |
|
---|
166 | private void DistributeReward(double reward) {
|
---|
167 | behaviourPolicy.UpdateReward(stateChain, reward);
|
---|
168 | greedyPolicy.UpdateReward(stateChain, reward);
|
---|
169 | }
|
---|
170 |
|
---|
171 |
|
---|
172 | private void Reset() {
|
---|
173 | behaviourPolicy.Reset();
|
---|
174 | greedyPolicy.Reset();
|
---|
175 | maxSearchDepth = 0;
|
---|
176 | bestQuality = 0.0;
|
---|
177 | tries = 0;
|
---|
178 | rootNode = new TreeNode(problem.Grammar.SentenceSymbol.ToString(), new ReadonlySequence("$"));
|
---|
179 | }
|
---|
180 |
|
---|
181 | public bool Done() {
|
---|
182 | int selectedStateIdx;
|
---|
183 | return !behaviourPolicy.TrySelect(random, rootNode.phrase, GenerateFollowStates(rootNode), out selectedStateIdx);
|
---|
184 | }
|
---|
185 |
|
---|
186 | #region introspection
|
---|
187 | public void PrintStats() {
|
---|
188 | Console.WriteLine("depth: {0,5} tries: {1,5} best phrase {2,50} bestQ {3:F3}", maxSearchDepth, tries, bestPhrase, bestQuality);
|
---|
189 |
|
---|
190 | // use behaviour strategy to generate the currently prefered sentence
|
---|
191 | var policy = behaviourPolicy;
|
---|
192 |
|
---|
193 | var n = rootNode;
|
---|
194 |
|
---|
195 | while (n != null) {
|
---|
196 | var phrase = n.phrase;
|
---|
197 | Console.ForegroundColor = ConsoleColor.White;
|
---|
198 | Console.WriteLine("{0,-30}", phrase);
|
---|
199 | var children = n.children;
|
---|
200 | if (children == null || !children.Any()) break;
|
---|
201 | var values = children.Select(ch => policy.GetValue(ch.phrase));
|
---|
202 | var maxValue = values.Max();
|
---|
203 | if (maxValue == 0) maxValue = 1.0;
|
---|
204 |
|
---|
205 | // write phrases
|
---|
206 | foreach (var ch in children) {
|
---|
207 | SetColorForValue(policy.GetValue(ch.phrase) / maxValue);
|
---|
208 | Console.Write(" {0,-4}", ch.phrase.Substring(Math.Max(0, ch.phrase.Length - 3), Math.Min(3, ch.phrase.Length)));
|
---|
209 | }
|
---|
210 | Console.WriteLine();
|
---|
211 |
|
---|
212 | // write values
|
---|
213 | foreach (var ch in children) {
|
---|
214 | SetColorForValue(policy.GetValue(ch.phrase) / maxValue);
|
---|
215 | Console.Write(" {0:F2}", policy.GetValue(ch.phrase) * 10.0);
|
---|
216 | }
|
---|
217 | Console.WriteLine();
|
---|
218 |
|
---|
219 | // write tries
|
---|
220 | foreach (var ch in children) {
|
---|
221 | SetColorForValue(policy.GetValue(ch.phrase) / maxValue);
|
---|
222 | Console.Write(" {0,4}", policy.GetTries(ch.phrase));
|
---|
223 | }
|
---|
224 | Console.WriteLine();
|
---|
225 | int selectedChildIdx;
|
---|
226 | if (!policy.TrySelect(random, phrase, children.Select(ch => ch.phrase), out selectedChildIdx)) {
|
---|
227 | break;
|
---|
228 | }
|
---|
229 | n = n.children[selectedChildIdx];
|
---|
230 | }
|
---|
231 |
|
---|
232 | Console.ForegroundColor = ConsoleColor.White;
|
---|
233 | Console.WriteLine("-------------------");
|
---|
234 | }
|
---|
235 |
|
---|
236 | private void SetColorForValue(double v) {
|
---|
237 | Console.ForegroundColor = ConsoleEx.ColorForValue(v);
|
---|
238 | }
|
---|
239 | #endregion
|
---|
240 |
|
---|
241 | }
|
---|
242 | }
|
---|