Dataset Viewer
Auto-converted to Parquet Duplicate
checkers
stringclasses
5 values
1
stringclasses
7 values
2
stringclasses
7 values
120
stringclasses
2 values
96.3
stringlengths
2
9
2.5
stringlengths
1
8
546.2
stringlengths
2
5
151
stringlengths
2
6
checkers
2
3
120
64.2
25
100.3
1297
checkers
3
4
120
57.1
20.8
49.1
4031
checkers
4
5
120
62.5
31.7
88
15645
checkers
5
6
120
62.9
35.8
91
49528
checkers
6
7
120
63.8
25.8
97.2
181183
chess
1
2
120
86.7
23.3
320.7
15200
chess
2
3
120
86.3
19.2
314.6
59662
chess
3
4
120
78.8
29.2
225.1
307045
connect4
1
2
120
100
0
952.8
30
connect4
2
3
120
51.3
15.8
8.6
231
connect4
3
4
120
53.8
7.5
25.9
566
connect4
4
5
120
57.9
12.5
55
1296
connect4
5
6
120
51.3
9.2
8.6
2537
connect4
6
7
120
57.9
7.5
55
5747
game
from_depth
to_depth
games
score_pct
draw_pct
elo
ms
othello
1
2
120
97.5
0
610.4
629
othello
2
3
120
71.7
3.3
159.7
4683
othello
3
4
120
79.2
0
229.4
16609
othello
4
5
120
73.3
0
174
73872
othello
5
6
120
65
1.7
106.6
270455

Negamax Search-Depth Scaling — What Is One Ply of Search Worth?

Measured strength gain, in Elo per extra ply of search depth, for four depth-limited negamax + alpha-beta board-game engines: Connect 4, Checkers, Othello, and Chess. Strength is defined as a fixed search(state, {maxDepth: N}) depth — a deterministic, hardware-independent knob — so every number reproduces from a seed.

Headline numbers (120 games/step, seed 20260819)

The single biggest gain is almost always the first extra ply — depth-1 is a near-random blunderer:

Game 1 → 2 ply (Elo) After the first ply
Connect 4 +952.8 collapses to near-noise (+9 to +55/ply, odd/even parity wobble)
Othello +610.4 bounces (no clean decay)
Checkers +546.2 plateaus around +90/ply
Chess +320.7 exception — 2→3 ply is nearly equal (+315); huge branching leaves plenty to discover even shallow

Files

  • data/elo-per-ply.csv — one row per adjacent depth step: game, from_depth, to_depth, games, score_pct, draw_pct, elo, ms.

Method

  • Strength = search depth. Deterministic search ⇒ hardware-independent, seed-reproducible results.
  • Because the search is deterministic, games are diversified with random openings played with both colours (each opening played twice, sides swapped, so first-move advantage cancels). Draws score ½.
  • Elo per step = 400 · log10(p / (1 − p)) from the colour-balanced score p.

Reproduce

Byte-identical engine snapshots and the harness are published alongside the study:

node negamax-harness.mjs --game othello --openings 60        # full ladder for one game
node negamax-harness.mjs --game chess --pair 2 --openings 60  # a single depth step (3 -> 4)

Citation

LK Forge (2026). What Is One Ply of Search Worth? — Negamax Search-Depth Scaling. https://lkforge.com/blog/search-depth-scaling/

Everything runs client-side, no login, no tracking. Full method and interactive charts on the write-up above.

Downloads last month
25