Skip to content

00-SERIES-OVERVIEW.md

Latest commit

 

History

History
55 lines (43 loc) · 3.89 KB

00-SERIES-OVERVIEW.md

File metadata and controls

55 lines (43 loc) · 3.89 KB
Raw

Titanium Engine — video series overview

Save this file. Record in any order; episodes are checkpointed in git.

Arc

  1. Scraped a Quoridor site — AI is on a server, not in the bundle
  2. Built Titanium Engine in Rust — rules first, speed second
  3. Perft caught our first real bug — fixed with divide debugging
  4. Plot twist: Rust crushes perft 3; perft 4 needed make/unmake + TT + flood fill — now ~3.4s (not minutes)
  5. Threading prep — Engine layout + root-parallel perft bench (thread-bench)
  6. Hybrid MCTS → pure αβ + adaptive LMR (MCTS deprecated, routes to negamax)
  7. Search hardening — negamax stop/tie-break/qsearch/CAT gap bugs; raw CAT overlay on web
  8. Perft closure — timed d4 regression test; incremental board masks tried and rejected
  9. Next: eval quality, BFS cache in search, opening depth vs Gorisanson

Episode list

# File Topic
01 01-path-bfs.md Bitboards + BFS
02 02-legal-moves.md 131 moves, JS parity
03 03-perft.md Divide harness
04 04-bench.md Criterion / NPS
05 05-first-perft-bug.md d8v lateral wall bug
06 06-threading-prep.md Titanium vs Titanium
07 07-ai-opponents.md Gorisanson local boss
08 08-greedy-ui-lab.md Testing lab UI + greedy genmove
10 10-hybrid-search.md MCTS↔minimax (historical; MCTS deprecated)
11 11-search-hardening.md Weird moves, gaps, qsearch, CAT UI
— 00-HOW-THE-ENGINE-WORKS.md Narration script
— CAT-VIEW-UI.md Debug overlay spec
— BUG-DIARY.md All plot twists
— PERFT-BENCHMARKS.md Depth timings

Git checkpoints

See README.md for branch names and commit hashes.

Future: every tagged checkpoint → build → round-robin → Elo ladder (TOURNAMENT-ROADMAP.md). Playable opponents start at episode 07 (αβ); earlier tags are perft/speed benchmarks only.

Competitors (reference only)

Project Role in our repo
gorisanson/quoridor-ai MCTS + heuristics; perft cross-check
pavlosdais/Quoridor C αβ competition engine; wall geometry
quoridor-ai.netlify.app Rules oracle (scraped)

Do NOT run in videos (outdated warnings)

  • perft 4 takes minutes — obsolete. Release build: ~3.4s, 247M nodes. Safe for demos on idle CPU.
  • Still prefer perft 3 or perft-race 3 for quick correctness beats (2M nodes, instant).
  • Do not run perft under heavy parallel cargo build — CPU contention inflates times (~10× false regression).
  • Use cargo test --release perft_depth4 -- --ignored --nocapture for timed regression with 10s cap.