Learning Without Backpropagation

A multiplayer Tetris sandbox where two gradient‑free learners improve through play:

• Genetic Algorithm (GA) — population‑based evolution of evaluation weights
• Markov Chain (MC) — table‑based transition learning from observed gameplay

No neural nets, no gradients, no backpropagation — just selection, mutation, crossover, and empirical transition counts.

Live demo: https://tetris-evolve.vercel.app
Quick start: clone → open index.html in any modern browser (no build step).

---

Why “No‑Backprop”?

This project demonstrates two classic, gradient‑free ways to make game‑playing agents improve:

• Genetic search treats the policy as a black box and evolves it by keeping better performers and randomly recombining/mutating their parameters.
• Markov learning builds a transition probability table directly from experience and uses it to bias future choices — an empirical, count‑based learner.

Both approaches avoid gradient descent and the chain rule entirely, yet they still get better with more play.

---

How Each Learner Improves

1) Genetic Algorithm (GA)

• Keep a population of candidate policies (“genomes”), each a vector of weights used to score potential moves.
• Let every genome play; compute a fitness (e.g., lines cleared / survival time / penalties for bad board shapes).
• Select higher‑fitness genomes, crossover pairs, and mutate weights (add small noise).
• Repeat for the next generation; persist the best genomes to localStorage so improvements carry across sessions.

Repo defaults: population ≈ 50 and mutation strength ≈ 0.1 (editable in code). The best‑performing weights are automatically saved, and there’s a dedicated population visualization page: ga_visualization.html.

Storage key for GA: tetris_ai_weights_v2_population (for persisted weight populations).

2) Markov Chain (MC)

• While games run, the MC agent observes sequences of (state, action, next‑state) and increments counts.
• It converts counts into transition probabilities, using them to bias future actions from similar states.
• When there’s no data for a situation, it falls back to a simple heuristic evaluation.
• Learned transition tables persist in the browser so play today informs play tomorrow.

Storage key for MC: tetris_markov_chains_v2_human_learned.

---

See It Learn (5‑Minute Demo)

1. Start from scratch (optional but recommended)
   Open the browser console and clear the saved learners:

   localStorage.removeItem('tetris_ai_weights_v2_population');
   localStorage.removeItem('tetris_markov_chains_v2_human_learned');