---

name: ds-online-experiments compatibility: opencode completeness: 95 content-types:

• code
• guidance
• do-dont
• examples description: '"Implements multi-armed bandits, contextual bandits, exploration-exploitation tradeoff, and online learning algorithms"' license: MIT maturity: stable metadata: domain: coding output-format: code related-skills: ds-ab-testing, ds-experimental-design, ds-metrics-and-kpis role: implementation scope: implementation triggers: multi-armed bandits, bandits, contextual bandits, exploration exploitation online learning archetypes:
  • tactical
  • generation anti_triggers:
  • brainstorming
  • vague ideation
  • code golf
  • over-engineering response_profile: verbosity: low directive_strength: high abstraction_level: operational version: "1.0.0"

---

Online Experiments

Comprehensive guide to online experiments in machine learning and data science workflows.

When to Use This Skill

• Solving real-world experimentation & a/b testing problems
• Building machine learning pipelines with online experiments
• Implementing best practices for online experiments
• Optimizing model performance using online experiments techniques
• Learning industry-standard approaches to online experiments

When NOT to Use This Skill

• When using pre-built libraries without understanding underlying concepts
• For toy problems that don't require online experiments rigor
• When domain expertise in specific problem requires different approach
• If your problem doesn't require the complexity this skill provides

Purpose and Key Concepts

Online Experiments is a critical component of the machine learning workflow. This skill covers:

1. Theoretical foundations — Mathematical principles and statistical concepts
2. Practical implementation — Working code examples and patterns
3. Common pitfalls — Mistakes to avoid and how to recover from them
4. Best practices — Industry-standard approaches and optimization techniques

Core Workflow

1. Understand the problem — Clearly define what you're solving for
2. Select approach — Choose the right technique for your data and constraints
3. Implement solution — Write clean, tested code following best practices
4. Validate results — Verify your implementation with tests and validation
5. Optimize performance — Improve efficiency and accuracy incrementally

Implementation Patterns

Pattern 1: Basic Online Experiments

import numpy as np
from typing import Dict, List, Tuple

class EpsilonGreedyBandit:
    """Basic epsilon-greedy multi-armed bandit implementation."""
    
    def __init__(self, n_arms: int, epsilon: float = 0.1) -> None:
        if n_arms <= 0:
            raise ValueError("n_arms must be a positive integer")
        self.n_arms = n_arms
        self.epsilon = max(0.0, min(1.0, epsilon))
        self.q_values = np.zeros(n_arms)
        self.arm_counts = np.zeros(n_arms)
        
    def select_arm(self) -> int:
        if np.random.random() < self.epsilon:
            return int(np.random.randint(self.n_arms))
        return int(np.argmax(self.q_values))
        
    def update(self, arm: int, reward: float) -> None:
        if not 0 <= arm < self.n_arms:
            raise ValueError(f"Arm index {arm} out of range [0, {self.n_arms})")
        self.arm_counts[arm] += 1
        n = self.arm_counts[arm]
        old_q = self.q_values[arm]
        self.q_values[arm] = old_q + (1.0 / n) * (reward - old_q)
        
    def get_statistics(self) -> Dict[str, List[float]]:
        return {
            'avg_reward': [float(np.mean(self.q_values))]
            'arm_counts': self.arm_counts.tolist()
            'q_values': self.q_values.tolist()
        }

Pattern 2: Production-Ready Online Experiments

import logging
import numpy as np
import pandas as pd
from typing import Any, Dict, List, Optional

logger = logging.getLogger(__name__)

class UCBBandit:
    """Upper Confidence Bound bandit for production online experiments."""
    
    def __init__(self, n_arms: int, confidence: float = 2.0) -> None:
        if n_arms <= 0:
            raise ValueError("n_arms must be positive")
        self.n_arms = n_arms
        self.confidence = confidence
        self.q_values = np.zeros(n_arms)
        self.arm_counts = np.zeros(n_arms)
        self.total_pulls = 0
        
    def select_arm(self) -> int:
        if self.total_pulls < self.n_arms:
            return int(np.argmin(self.arm_counts))
        ucb_values = self.q_values + self.confidence * np.sqrt(
            np.log(self.total_pulls) / self.arm_counts
        )
        return int(np.argmax(ucb_values))
        
    def update(self, arm: int, reward: float) -> None:
        self.arm_counts[arm] += 1
        self.total_pulls += 1
        n = self.arm_counts[arm]
        old_q = self.q_values[arm]
        self.q_values[arm] = old_q + (1.0 / n) * (reward - old_q)
        
    def execute(self, data: pd.DataFrame) -> Dict[str, Any]:
        required_cols = {'reward', 'arm'}
        if not required_cols.issubset(data.columns):
            raise ValueError(f"DataFrame must contain columns: {required_cols}")
            
        results: List[Dict[str, Any]] = []
        for _, row in data.iterrows():
            arm = int(row['arm'])
            reward = float(row['reward'])
            self.update(arm, reward)
            results.append({
                'arm': arm, 
                'reward': reward, 
                'q_estimate': float(self.q_values[arm])
            })
            
        logger.info(f"Processed {len(results)} observations across {self.n_arms} arms")
        return {
            'final_q_values': self.q_values.tolist()
            'arm_counts': self.arm_counts.tolist()
            'total_pulls': self.total_pulls
            'history': results
        }

Pattern 3: BAD vs GOOD Implementation

# BAD: Hardcoded values, no type hints, bypasses validation, monolithic function
def bad_bandit(data):
    q = [0.0, 0.0, 0.0]
    for i in range(len(data)):
        if i < 3:
            arm = i
        else:
            arm = q.index(max(q))
        q[arm] = q[arm] + 0.1 * (data[i] - q[arm])
    return q

# GOOD: Parameterized, typed, validated, decomposed, follows SOLID principles
import numpy as np
from typing import List, Dict

def good_bandit(rewards: List[float], n_arms: int = 3, alpha: float = 0.1) -> Dict[str, List[float]]:
    if n_arms <= 0 or not rewards:
        raise ValueError("Invalid parameters")
        
    q_estimates = np.zeros(n_arms)
    arm_counts = np.zeros(n_arms)
    
    for t, reward in enumerate(rewards):
        if t < n_arms:
            arm = t
        else:
            arm = int(np.argmax(q_estimates))
            
        arm_counts[arm] += 1
        n = arm_counts[arm]
        q_estimates[arm] += alpha * (reward - q_estimates[arm])
        
    return {
        'estimated_means': q_estimates.tolist()
        'pull_counts': arm_counts.tolist()
    }

Best Practices

• ✅ Always validate your implementation on test data
• ✅ Document your assumptions and methodology
• ✅ Use version control for reproducibility
• ✅ Monitor performance metrics in production
• ✅ Periodically review and update your approach
• ✅ Test with edge cases and outliers
• ✅ Log all significant operations for debugging
• ✅ Follow SOLID principles for class design and DRY for utility functions

Common Pitfalls

| Pitfall | Problem | Solution | |

---

---

Constraints

MUST DO

• Validate all data preprocessing steps are fit-only on training data, never on validation or test sets
• Implement reproducible pipelines with fixed random seeds and deterministic operations where possible
• Report model performance with confidence intervals via bootstrapping or cross-validation across multiple runs
• Log all experiments with parameters, metrics, and artifacts using MLflow or equivalent tracking system

MUST NOT DO

• Do not evaluate a model on the same data used for training — always hold out a proper test set
• Avoid overfitting to the validation set by limiting hyperparameter search iterations
• Never use features that can only be computed at inference time (look-ahead bias)
• Do not report single-run accuracy without statistical significance testing or error bars

Live References

Authoritative documentation links for this skill's domain. The model follows markdown links at load time to resolve external references and inline content.

• Online Controlled Experiments (Microsoft Research)
• Optimizely Online Experimentation Guide
• VWO A/B Testing Best Practices
• Experimentation at Scale — Netflix Eng Blog
• Google Optimize Documentation