learning_path_2_intermediate.md

🏃‍♂️ AI4R Learning Path #2 – Intermediate Track

From “I can run a model” to “I can tune, extend, and build smart stuff that actually works.”

You’ve trained your first models, tweaked some clusters, maybe even backprop’ed a neural net. Nice.

Now it’s time to go deeper: tune ensembles, shape rewards, mess with distance functions, evolve a Sudoku solver—and crack open a Transformer like it’s no big deal. Let’s go.

🛠️ Prerequisites

• You’ve finished the Beginner Track (or already speak fluent require 'ai4r')
• You’ve cloned the repo and can run tests with bundle exec rake
• You’re cool with Git (or brave enough to fake it)

🌳 Module 1 – Forests, Boosting & the Art of Tuning

🧠 Learn: tree ensembles are overpowered—and totally tunable.

1. Read through random_forest.rb and gradient_boosting.rb.
2. Copy compare_all.rb to forest_vs_boost.rb.
3. Play with num_trees, max_depth, learning_rate, etc.
4. Graph accuracy vs. hyperparameters—bonus points for pretty colors.

✅ You now understand:

• Why RandomForest works out of the box
• Why GradientBoosting can beat it—with tuning

🔬 Module 2 – Distance is a Choice, Not a Fact

🧠 Learn: the way your model measures similarity changes everything.

1. Add chebyshev to lib/ai4r/proximity.rb
2. Swap it into KMeans or DBSCAN in kmeans_vs_dbscan.rb
3. Watch clusters shift dramatically—even though the data didn’t

✅ You now understand:

• How distance metrics control clustering
• That Euclidean isn't always right

🧬 Module 3 – Roll Your Own Genetic Algorithm

🧠 Learn: if you can encode a problem, you can evolve a solution.

1. Peek at tsp_chromosome.rb.
2. Build a new chromosome (e.g., 4x4 Sudoku)
3. Plug it into the GA engine and evolve till solved
4. Mutation rate too high? Watch it crash. Too low? It stalls.

✅ You now understand:

• Chromosome design = creative encoding
• Mutation & crossover = controlled chaos

🎮 Module 4 – Q-Learning, but Make It Smarter

🧠 Learn: shaping rewards teaches agents faster than yelling “wrong!”

1. Start with the grid-world in docs/reinforcement_learning.md
2. Add:
   • Negative reward for each step
   • Big bonus for early success
3. Compare convergence speed (how many episodes to learn)

✅ You now understand:

• Rewards aren’t just feedback—they’re a strategy
• Agents learn what you incentivize, not what you want

🤖 Module 5 – A Tiny Transformer That You Can Actually Read

🧠 Learn: break down attention, multi-heads, and sequence scaling—line by line.

1. Open transformer.rb—yes, read the whole thing
2. Copy the inline example to mini_transformer.rb
3. Tweak:
   • Sequence length (10 vs 50)
   • Head count (2 vs 4)
4. Watch parameter count, runtime, and output patterns shift

✅ You now understand:

• Why attention is expensive
• What multi-head actually means in code

📊 Module 6 – Metrics That Matter

🧠 Learn: building your own metric is the most honest way to evaluate.

1. Check out bench/common/metrics.rb
2. Add a balanced_accuracy or f_beta function
3. Re-run a classifier bench and print your custom metric
4. Push it to your fork—you’re officially contributing

✅ You now understand:

• Metrics shape how models are judged
• “Accuracy” alone isn’t enough

🏁 Capstone – Mix & Match Time

Pick a real dataset (UCI, Kaggle, or scrape your own). Then:

• Train a boosted classifier
• Cluster users or records
• Recommend an action policy using Q-Learning
• Tune, measure, iterate—then ship your notebook or .rb file

Document your decisions like this:

1. What you’re solving
2. What you tried
3. What worked (and didn’t)
4. What you learned

✅ You’ve combined multiple AI techniques into a single project. Boom.

🥈 You’ve Leveled Up

You now:

• Know how to tune tree-based models
• Customize core behavior (metrics, distances, fitness)
• Build your own problem encodings
• Understand RL reward design
• Actually get Transformers
• Can extend and contribute to AI4R itself

Next stop: Advanced Track — Hidden Markov Models, Monte Carlo Tree Search, custom benchmarks, and serious optimization.

Ready to break things?