← Experiments

Leaderboard — Elo & head-to-head win ratios

Round-robin at starting authority 10, 6 games per pair · the data lesson_seven.py prints, on a page

Every policy plays every other in a round robin; Elo starts at 1500, K=32. The matrix shows each row method's win rate against each column method.

Elo ratings

MethodElo
V4 Ordered List (Play coins first)1669.7
AlphaStar (under-trained, 140g)1626.9
V2 Ordered List1591.7
AlphaStar (well-trained, 500g)1528.0
V1 Choose randomly, excluding end turn1523.3
V3 Ordered List (cautious scraping)1333.0
V0 Choose completely randomly1227.3

Win-ratio matrix (row vs. column)

V0 Choose completely randomlyV1 Choose randomly, excluding end turnV2 Ordered ListV3 Ordered List (cautious scraping)V4 Ordered List (Play coins first)AlphaStar (under-trained, 140g)AlphaStar (well-trained, 500g)
V0 Choose completely randomly0%0%0%0%0%17%
V1 Choose randomly, excluding end turn100%33%100%33%50%33%
V2 Ordered List100%67%100%33%33%67%
V3 Ordered List (cautious scraping)100%0%0%17%17%0%
V4 Ordered List (Play coins first)100%67%67%83%83%33%
AlphaStar (under-trained, 140g)100%50%67%83%17%67%
AlphaStar (well-trained, 500g)83%67%33%100%67%33%

The two AlphaStar entries are the same architecture at two training depths, included to show Elo rising with training. See the policies page for what each method is.

How the fully-trained agents' games end

A game can end three ways: decisively (a player's authority is driven to 0), by the engine's 50-turn cap (whoever has more authority wins), or by our 200-action move cap (a safety bound we impose so a single game can't run away). We halved the move cap from 400 to 200; the counts below — from 30 games of the well-trained models head-to-head at each authority — check we did not overdo it.

Starting authorityDecisive (lethal)Turn cap (50 turns)Move cap (200 actions)
10 authority30 (100%)0 (0%)0 (0%)
50 authority1 (3%)25 (83%)4 (13%)

4 game(s) hit the 200-action move cap — the cap is ending games that had not yet resolved, so it may be set too low. See the engineering challenge for why high-authority games are long and expensive.

← All experiments · Home