classic_bandit 0.1.0 → 0.1.2

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: 9ce78adca73218db90a909daec3ccbb457417620d78ae2566c0a32ac7024a313
-  data.tar.gz: 9b6debd514866747242bd00de9db4d97d1fd95e5a03166e52c96b812daa3d7f6
+  metadata.gz: 9e7b77e4f21e909a3515df0c5785b2f321ce3cf2977d825f0190adcd9f15a82d
+  data.tar.gz: daeb7c6386d836256e0a32c65fc72e68ad02ad2283b4a6e89b8a4eb9eebf35ab
 SHA512:
-  metadata.gz: 80cc50641016e81853f766645d1fe9baf3c7bf4b09c8bc9303ce083dac1bec574d78432dd174953785e3dd37032e34c2248d22f47acd1c08c02078f02ad6f8a7
-  data.tar.gz: e5b534ad79cc2a91a95b617c315335cb8c399c7241308750d868b4eb13e60016382765ff1e83447495676414e1213a6b7212d31e9b18f2859063d5f491f80b25
+  metadata.gz: 7ca36345f798cc2e869cba8e0c3e41b33892771227fcd12c85c2d9973d6fb697a79b22366f6f59a7bb041772e1d4688147162b78f0d2530ec9d4dd9da5e45981
+  data.tar.gz: 1bea576eeb3b00795b40206279376fd9b1da678cfd9ce5157829d1be8983fcb7fa7f488d91733715d263e772de250eb1a79145790152e99cb771515453db3fbe

data/.rubocop.yml

@@ -1,5 +1,9 @@
 AllCops:
   TargetRubyVersion: 3.0
+  NewCops: disable
+  Include:
+    - 'lib/**/*'
+    - 'spec/**/*'
 
 Style/StringLiterals:
   EnforcedStyle: double_quotes
@@ -13,4 +17,7 @@ Metrics/BlockLength:
 
 Style/Documentation:
   Exclude:
-    - '**/*'
+    - '**/*'
+
+Metrics/MethodLength:
+  Max: 30

data/CLAUDE.md

@@ -0,0 +1,45 @@
+# CLAUDE.md
+
+This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.
+
+## Development Commands
+
+- **Run tests**: `bundle exec rspec`
+- **Lint code**: `bundle exec rubocop`
+- **Run all checks**: `bundle exec rake` (runs both tests and linting)
+- **Install dependencies**: `bundle install`
+
+The project uses RuboCop for linting with explicit config file reference in Rakefile:15.
+
+## Architecture Overview
+
+ClassicBandit is a Ruby gem implementing multi-armed bandit algorithms for A/B testing and optimization. The library uses Zeitwerk for autoloading and follows a modular design pattern.
+
+### Core Components
+
+**Arm (`lib/classic_bandit/arm.rb`)**: Represents a bandit arm with trial/success tracking and mean reward calculation.
+
+**ArmUpdatable (`lib/classic_bandit/arm_updatable.rb`)**: Shared module providing `update(arm, reward)` method for all bandit algorithms. Validates rewards are 0 or 1.
+
+**Algorithm Implementations**:
+- **EpsilonGreedy**: Simple ε-greedy with exploration/exploitation balance
+- **UCB1**: Upper Confidence Bound without explicit parameters  
+- **Softmax**: Temperature-based Boltzmann distribution selection
+- **ThompsonSampling**: Bayesian approach with Beta-Bernoulli model using custom Gamma random number generation
+
+### Key Design Patterns
+
+All algorithms implement the same interface:
+- `select_arm()` → returns an Arm instance
+- `update(arm, reward)` → updates arm statistics (from ArmUpdatable)
+- Handle untested arms by random selection
+
+The Thompson Sampling implementation includes custom statistical functions (`gamma_random`, `normal_random`) using Marsaglia-Tsang method for Gamma distribution sampling.
+
+### Testing Structure
+
+Tests are organized in `spec/` with individual algorithm specs and a main gem spec. Uses standard RSpec testing framework.
+
+### Example Usage
+
+The `example/` directory contains simulation scripts demonstrating algorithm comparison with Gnuplot visualization, showing realistic usage patterns with pre-populated arm statistics.

data/Rakefile

@@ -7,6 +7,8 @@ RSpec::Core::RakeTask.new(:spec)
 
 require "rubocop/rake_task"
 
-RuboCop::RakeTask.new
+RuboCop::RakeTask.new do |task|
+  task.options = ["--config", ".rubocop.yml"] # 明示的に設定ファイルを指定
+end
 
 task default: %i[spec rubocop]

data/example/Gemfile

@@ -0,0 +1,12 @@
+# frozen_string_literal: true
+
+source "https://rubygems.org"
+
+# gem "rails"
+
+gem "gnuplot", "~> 2.6"
+gem "matrix", "~> 0.4.2"
+
+gem "zeitwerk", "~> 2.7"
+
+gem "classic_bandit", "~> 0.1.0"

data/example/Gemfile.lock

@@ -0,0 +1,20 @@
+GEM
+  remote: https://rubygems.org/
+  specs:
+    classic_bandit (0.1.0)
+    gnuplot (2.6.2)
+    matrix (0.4.2)
+    zeitwerk (2.7.1)
+
+PLATFORMS
+  ruby
+  x86_64-linux
+
+DEPENDENCIES
+  classic_bandit (~> 0.1.0)
+  gnuplot (~> 2.6)
+  matrix (~> 0.4.2)
+  zeitwerk (~> 2.7)
+
+BUNDLED WITH
+   2.6.2

data/example/beta_random.rb

@@ -0,0 +1,87 @@
+# frozen_string_literal: true
+
+require "gnuplot"
+
+def gamma_random(alpha) # rubocop:disable Metrics/AbcSize
+  return gamma_random(alpha + 1) * rand**(1.0 / alpha) if alpha < 1
+
+  d = alpha - 1.0 / 3
+  c = 1.0 / Math.sqrt(9 * d)
+
+  loop do
+    x = normal_random
+    v = (1 + c * x)**3
+    
+    next if v <= 0
+    
+    u = rand
+    
+    # Squeeze test
+    return d * v if u < 1 - 0.0331 * x**4
+    
+    # Full test
+    return d * v if Math.log(u) < 0.5 * x * x + d - d * v + d * Math.log(v)
+  end
+end
+
+def normal_random
+  r = Math.sqrt(-2 * Math.log(rand))
+  theta = 2 * Math::PI * rand
+  r * Math.cos(theta)
+end
+
+def beta_function(alpha, beta)
+  gamma_alpha = Math.lgamma(alpha)[0]
+  gamma_beta = Math.lgamma(beta)[0]
+  gamma_apb = Math.lgamma(alpha + beta)[0]
+  Math.exp(gamma_alpha + gamma_beta - gamma_apb)
+end
+
+def beta_pdf(x, alpha, beta)
+  return 0 if x <= 0 || x >= 1
+
+  x**(alpha - 1) * (1 - x)**(beta - 1) / beta_function(alpha, beta)
+end
+
+data = Array.new(10_000) do
+  x1 = gamma_random(41)
+  x2 = gamma_random(61)
+  x1 / (x1 + x2)
+end
+
+Gnuplot.open do |gp|
+  Gnuplot::Plot.new(gp) do |plot|
+    plot.title  "Beta distribution histogram"
+    plot.xlabel "Value"
+    plot.ylabel "Frequency"
+
+    min_val = 0.0
+    max_val = 1.0
+    bin_count = 60.0
+    bin_width = (max_val - min_val) / bin_count
+
+    plot.xrange "[0:1]"
+    total_count = data.length.to_f
+
+    plot.set "style data histograms"
+    plot.set "style fill solid 0.5"
+
+    bins = Hash.new(0)
+    bin_count.to_i.times.each { |i| bins[i * bin_width] = 0 }
+    data.each { |v| bins[(v / bin_width).floor * bin_width] += 1 }
+    bins.transform_values! { |v| v / (total_count * bin_width) }
+
+    plot.data << Gnuplot::DataSet.new([bins.keys, bins.values]) do |ds|
+      ds.with = "boxes"
+      ds.title = "Empirical"
+    end
+
+    x_points = (0..100).map { |i| i / 100.0 }
+    y_points = x_points.map { |x| beta_pdf(x, 41, 61) }
+    plot.data << Gnuplot::DataSet.new([x_points, y_points]) do |ds|
+      ds.with = "lines"
+      ds.linewidth = 2
+      ds.title = "Theoretical PDF"
+    end
+  end
+end

data/example/simulation.rb

@@ -0,0 +1,73 @@
+# frozen_string_literal: true
+
+require 'classic_bandit'
+require 'gnuplot'
+
+bandits = {
+  "UCB1" => ClassicBandit::Ucb1.new(arms: [
+      ClassicBandit::Arm.new(id: 0, trials: 1000, successes: 120),
+      ClassicBandit::Arm.new(id: 1, trials: 1000, successes: 110),
+      ClassicBandit::Arm.new(id: 2, trials: 1000, successes: 100),
+  ]),
+  "Thompson sampling" => ClassicBandit::ThompsonSampling.new(arms: [
+      ClassicBandit::Arm.new(id: 0, trials: 1000, successes: 120),
+      ClassicBandit::Arm.new(id: 1, trials: 1000, successes: 110),
+      ClassicBandit::Arm.new(id: 2, trials: 1000, successes: 100),
+  ])
+}
+
+arm0_counts = Hash.new(0)
+arm0_probs = {}
+bandits.keys.each { |key| arm0_probs[key] = [] }
+x_values = []
+
+10000.times.each do |i|
+  bandits.each do |key, bandit|
+    # 最初の500回はランダム
+    if i < 500
+      arm = bandit.arms.sample
+    else
+      arm = bandit.select_arm
+    end
+    reward = rand <= arm.mean_reward ? 1 : 0
+    bandit.update(arm, reward)
+
+    if arm.id == 0
+      arm0_counts[key] += 1
+    end
+
+    arm0_prob = arm0_counts[key].to_f / (i + 1)
+    arm0_probs[key] << arm0_prob
+  end
+
+  x_values << i + 1
+end
+
+Gnuplot.open do |gp|
+  Gnuplot::Plot.new(gp) do |plot|
+    plot.title  "Bandit Selection Probability"
+    plot.xlabel "Iterations"
+    plot.ylabel "Probability"
+    
+    # y軸の範囲を0-1に設定
+    plot.yrange "[0:1]"
+    
+    # グリッドを表示
+    plot.set "grid"
+    
+    # 線のスタイルを設定
+    plot.set "style line 1 linecolor rgb '#0060ad' linewidth 2"
+    plot.set "style line 2 linecolor rgb '#dd181f' linewidth 2"
+    
+    # 各アルゴリズムのデータをプロット
+    colors = ["#0060ad", "#dd181f"]
+    bandits.each_with_index do |(key, _), index|
+      plot.data << Gnuplot::DataSet.new([x_values, arm0_probs[key]]) do |ds|
+        ds.with = "lines"
+        ds.linewidth = 2
+        ds.linecolor = "rgb '#{colors[index]}'"
+        ds.title = key.to_s
+      end
+    end
+  end
+end

data/lib/classic_bandit/thompson_sampling.rb

@@ -14,26 +14,39 @@ module ClassicBandit
   class ThompsonSampling
     include ArmUpdatable
 
-    attr_reader :arms
+    attr_reader :arms, :alpha_prior, :beta_prior
+
+    # @param arms [Array<Arm>] Array of arms to choose from
+    # @param alpha_prior [Float] Prior parameter for successes (default: 1.0)
+    # @param beta_prior [Float] Prior parameter for failures (default: 1.0)
+    def initialize(arms:, alpha_prior: 1.0, beta_prior: 1.0)
+      raise ArgumentError, "alpha_prior must be positive" unless alpha_prior.positive?
+      raise ArgumentError, "beta_prior must be positive" unless beta_prior.positive?
 
-    def initialize(arms:)
       @arms = arms
+      @alpha_prior = alpha_prior
+      @beta_prior = beta_prior
     end
 
     def select_arm
-      return @arms.sample if @arms.all? { |arm| arm.trials.zero? }
-
       @arms.max_by { |arm| ts_score(arm) }
     end
 
     private
 
     def ts_score(arm)
-      return 0.0 if arm.trials.zero?
-      return 1.0 if arm.successes == arm.trials
+      alpha = arm.successes + @alpha_prior
+      beta = (arm.trials - arm.successes) + @beta_prior
+
+      beta_sample(alpha, beta)
+    end
 
-      x = gamma_random(arm.successes + 1)
-      y = gamma_random(arm.trials - arm.successes + 1)
+    def beta_sample(alpha, beta)
+      # Beta(1,1) = Uniform(0,1)
+      return rand if alpha == 1.0 && beta == 1.0 # rubocop:disable Lint/FloatComparison
+
+      x = gamma_random(alpha)
+      y = gamma_random(beta)
       x / (x + y)
     end
 
@@ -45,11 +58,18 @@ module ClassicBandit
       c = 1.0 / Math.sqrt(9 * d)
 
       loop do
-        z = normal_random
-        v = (1 + c * z)**3
+        x = normal_random
+        v = (1 + c * x)**3
+
+        next if v <= 0
+
         u = rand
 
-        return d * v if z > -1.0 / c && Math.log(u) < 0.5 * z * z + d * (1 - v + Math.log(v))
+        # Squeeze test
+        return d * v if u < 1 - 0.0331 * x**4
+
+        # Full test
+        return d * v if Math.log(u) < 0.5 * x * x + d - d * v + d * Math.log(v)
       end
     end
 

data/lib/classic_bandit/version.rb

@@ -1,5 +1,5 @@
 # frozen_string_literal: true
 
 module ClassicBandit
-  VERSION = "0.1.0"
+  VERSION = "0.1.2"
 end

metadata

@@ -1,14 +1,14 @@
 --- !ruby/object:Gem::Specification
 name: classic_bandit
 version: !ruby/object:Gem::Version
-  version: 0.1.0
+  version: 0.1.2
 platform: ruby
 authors:
 - Kohei Tsuyuki
-autorequire: 
+autorequire:
 bindir: exe
 cert_chain: []
-date: 2024-12-27 00:00:00.000000000 Z
+date: 2025-11-09 00:00:00.000000000 Z
 dependencies: []
 description: Implementation of classic multi-armed bandit algorithms in Ruby. Supports
   Thompson Sampling, UCB1, and Epsilon-Greedy strategies with a simple, consistent
@@ -23,9 +23,14 @@ files:
 - ".rspec"
 - ".rubocop.yml"
 - CHANGELOG.md
+- CLAUDE.md
 - LICENSE.txt
 - README.md
 - Rakefile
+- example/Gemfile
+- example/Gemfile.lock
+- example/beta_random.rb
+- example/simulation.rb
 - lib/classic_bandit.rb
 - lib/classic_bandit/arm.rb
 - lib/classic_bandit/arm_updatable.rb
@@ -42,7 +47,7 @@ metadata:
   homepage_uri: https://github.com/t-chov/classic_bandit
   source_code_uri: https://github.com/t-chov/classic_bandit
   changelog_uri: https://github.com/t-chov/classic_bandit/blob/main/CHANGELOG.md
-post_install_message: 
+post_install_message:
 rdoc_options: []
 require_paths:
 - lib
@@ -58,7 +63,7 @@ required_rubygems_version: !ruby/object:Gem::Requirement
       version: '0'
 requirements: []
 rubygems_version: 3.2.33
-signing_key: 
+signing_key:
 specification_version: 4
 summary: A Ruby library for classic (non-contextual) multi-armed bandit algorithms
 test_files: []