Introduction to Machine Learning
Machine Learning (ML) is the engine that powers most modern AI agents. It allows agents to learn from data rather than being explicitly programmed.
Why Agents Need ML
- Environments are too complex for hand-coded rules
- Patterns in data are too subtle for humans to specify
- Agents need to adapt to changing conditions
- ML enables generalization from examples
Supervised Learning
Learn from labeled examples (input → correct output).
- Classification: Predict a category (spam/not spam, cat/dog)
- Regression: Predict a number (house price, temperature)
# Simple supervised learning example
from sklearn.ensemble import RandomForestClassifier
# Training data: features → labels
X_train = [[0, 0], [1, 1], [2, 2], [3, 3]]
y_train = [0, 0, 1, 1]
# Train model
model = RandomForestClassifier()
model.fit(X_train, y_train)
# Predict
prediction = model.predict([[1.5, 1.5]]) # → 0 or 1
Unsupervised Learning
Find patterns in data without labels.
- Clustering: Group similar items (customer segments)
- Dimensionality reduction: Compress data while preserving structure
- Anomaly detection: Find unusual patterns
Reinforcement Learning
Learn by trial and error with rewards and penalties.
- Agent takes actions in an environment
- Receives rewards or penalties
- Learns a policy that maximizes cumulative reward
- Most relevant to AI agent development!
# Reinforcement learning concept
# Agent learns: state → action mapping (policy)
# Goal: maximize total reward over time
# Q-learning update rule:
# Q(s, a) = Q(s, a) + α * (reward + γ * max(Q(s', a')) - Q(s, a))
When to Use Which
| Paradigm | Use When | Agent Example |
| Supervised | Have labeled data | Spam filter agent |
| Unsupervised | Need to find structure | Customer segmentation agent |
| Reinforcement | Agent interacts with environment | Game-playing agent |
Key Takeaways
- ML enables agents to learn from experience rather than following fixed rules
- Supervised learning needs labeled data; unsupervised finds patterns; RL learns from rewards
- Reinforcement learning is the most natural fit for agent development
- Modern agents often combine multiple ML paradigms