turnip-analysis.md

I made a turnip-counter! It reads Super Smash Bros Melee replay data, searches for items pulled by Peach, and records information to a sqlite3 database!

The code looks specifically for items initially owned by me, so the good luck of other Peach players isn't counted:

/// Tokio task to send Peach item records to database worker.
async fn record_items(tx: Sender<DBCommand>, items: ItemLog, game_id: i64, me: Port) -> App<()> {
    for (item_id, history) in items {
        // only record my turnips
        if history.data.owner != me {
            continue;
        }
        // ...
    }
}

I'm using the peppi Rust library to parse replays. I used tokio for async, thiserror for error handling, sqlx for typed queries + db migrations, and walkdir for recursively walking through replay directories.

When Peach pulls items in melee, the items pulled are random, and we have expected values based on our understanding of Melee's RNG:

| Type                                   | Damage | Probability         |
|----------------------------------------|--------|---------------------|
| Normal                                 | 6-10%  | 4445/7424 ≈ 59.873% |
| Carrot Eyes                            | 6-10%  | 762/7424 ≈ 10.264%  |
| Line Eyes                              | 6-10%  | 635/7424 ≈ 8.553%   |
| Circle Eyes                            | 6-10%  | 381/7424 ≈ 5.132%   |
| Eyebrow Eyes                           | 6-10%  | 381/7424 ≈ 5.132%   |
| Wink                                   | 10-14% | 508/7424 ≈ 6.843%   |
| Dot Eyes                               | 16-20% | 127/7424 ≈ 1.711%   |
| Stitch Face                            | 34-38% | 127/7424 ≈ 1.711%   |
| [Mr. Saturn](/Mr._Saturn "Mr. Saturn") | —      | 3/768 ≈ 0.391%      |
| [Bob-omb](/Bob-omb "Bob-omb")          | —      | 2/768 ≈ 0.260%      |
| [Beam Sword](/Beam_Sword "Beam Sword") | —      | 1/768 ≈ 0.130%      |

With that information, I can compare my actual frequency of items pulled vs the predicted frequency of items pulled. It's very close! I have 418 replays from this month (October 2022) to analyze:

❯ DATABASE_URL=sqlite://turnips.db cargo sqlx database create
❯ DATABASE_URL=sqlite://turnips.db cargo sqlx migrate run --source db/migrations
Applied 0/migrate init (3.274057ms)
❯ DATABASE_URL=sqlite://turnips.db cargo run -- ~/Slippi/2022-10/
...

sqlite> SELECT COUNT(*) FROM items;
3505
sqlite> SELECT COUNT(*) FROM games;
418
sqlite> SELECT kind, CAST(COUNT(*) AS REAL) / (SELECT COUNT(*) FROM items) FROM items GROUP BY kind;
Beamsword      0.00114122681883024 -- ~0.11% actual vs 0.13% predicted
Bobomb         0.00228245363766049 -- ~0.23% actual vs 0.26% predicted
DotEyesTurnip  0.0145506419400856  -- ~1.46% actual vs 1.71% predicted
MrSaturn       0.00485021398002853 -- ~0.46% actual vs 0.39% predicted
NormalTurnip   0.888445078459344   -- ~88.84% actual vs 88.95% predicted (59.873 + 10.264 + 8.553 + 5.132 + 5.132)
StitchTurnip   0.0182596291012839  -- ~1.82% actual vs 1.71% predicted
WinkyTurnip    0.0704707560627675  -- ~7.04% actual vs 6.84% predicted