Where the project's ideas came from and what became of them. Many are the user's; others come from AlphaStar and OpenAI Five (what those papers found), from LLM reinforcement learning, from image generation, or from our own runs. Each experiment's numbers are in the experiment log.
Open ideas. Next experiments has the details; the base is fgself-238.
| idea | from | what it is | status |
|---|---|---|---|
| Pressure on the AI's economy | audit 20 (2026-11-18: long wins are games where the AI's economy broke) | the shaping's material counts the opponent's workers several times their cost: a reward, not a lesson | open (needs a small flag) |
| Revival under the cap | audit 20 | --revive-lesson 1 with the revival orders capped by --teach-floor (uncapped it was part of fgself-157's loss) |
open (needs a floor group) |
| Style specialists in the league; distillation behind a style token | ours (2026-11-17, the user: bolder changes for strength and diversity) | specialists pushed into one style each (first: air) as fixed league members, then one policy that plays each style on demand | queued |
| Stacked capped lessons | ours (2026-11-17) | the spell, upgrade, army and expansion lessons together under the cap (one at a time: spells kept, army and upgrades level, the expansion -11 through its uncapped camp lesson) | queued (the flags exist) |
| Fine points around the unit | the user (2026-10-25) | a sub-cell within the chosen point bin (--fine-points 7) |
merged, not yet in a run |
| Fewer decisions, spatial point heads | AlphaStar (action rate), SC2LE | a "keep this order k steps" head; points on a map, not x then y | later |
| An APM cap | ours | every unit acts every half second now | before games against people |
| idea | from | result |
|---|---|---|
| The first hero weighted | audit 20 (2026-11-18: the first hero decides orc and human games) | --first-hero-weights Ofar:4,Hamg:4 (fgself-236): kept; its end 0.845 against insane over two bursts (orc 0.75 against 0.52), its average 0.955 in burst-237 |
| Race weights after the audit | audit 20 (orc the weakest race, the night elf AI the hardest) | --learner-races human:0.25,undead:0.25,orc:0.3,nightelf:0.2 (fgself-237): nothing once the first hero is weighted; not kept |
| A smaller learning rate | ours (2026-11-18: checkpoints swing, a smaller step may hold a higher level) | --lr 2e-5 (fgself-234): level; lr stays 3e-5 |
| Masks for orders the game refuses | bug finder 10 (2026-11-16) | spells' targets by the game's rules, toggles by state, learns only for heroes (--order-masks 7, fgself-231): kept, level strength, ~20% fewer refused orders |
| Late starts from lost positions | ours (2026-11-17, the bolder changes) | 2,912 states of games the learner went on to lose (fgself-230): more tech, no gain; not kept |
| Capped army and upgrade lessons | audit 16's lessons, under the cap (2026-11-17) | --army-lesson 1 (fgself-226): rare, level; --upgrade-lesson 1 (fgself-227): research up in training, level; not kept |
| The heroes' spells, capped | the user (2026-11-11: heroes rarely cast) | --spell-lesson 0.25 --teach-floor 0.005 (fgself-224): kept, ~17 taught spells a game, head to head 0.59-0.62 against its control |
| Expansions, gentler | ours (audit 15, a bug finder 2026-11-10) | a late expansion lesson (main mine under 30%) with the halls capped (fgself-223): expanded in 4% of games, -11; its camp lesson cannot be capped. Not kept |
| The z-modes' deadlines for the real map | bug finder 9 (2026-11-12) | --z-mode-time 2 (fgself-222): a third to a sixth of the violations, no gain; not kept |
| Weight averaging (SWA) | ours (2026-11-16: a checkpoint's level swings ~8 points, burst-216/220) | the mean of a stretch of checkpoints (scripts/average_checkpoints.py): the mean of 212's 81 at its best single point (0.845); in use, the bases since fgself-221 mostly start from one. One broke a race (fgself-236's 17-27M: night elf 0.38): averages are read by race |
| Unit attacks | the user (2026-11-12: workers survive attack-moves) | --attack-lesson (fgself-218 collapsed); with --teach-floor 0.05 (-219) it spread, at 0.002 (-220) -8 against its control: closed for now |
| A KL anchor cloned from people | AlphaStar (its KL to the supervised human policy) | people's replays again with the obs-full events, a clone from fgself-212: as the only anchor (fgself-215) h2h 0.31, fgself-100's pull swapped for the clone's; as a second anchor (--ref2, fgself-216) -5 against a null fork, narrower. Not kept |
| Late starts from the late game, in the real game | ours (2026-11-12: 102 of fgself-201's 118 losses past 18 minutes) |
--real-agents 0.5 --real-late-minutes 16,28 --real-late-from 16 (fgself-214): worse in the burst (0.69 against insane; human and undead fell); not kept |
| Faster, less noisy verdicts | the user (2026-11-13) | training games as a read (scripts/train_read.py, drifts between days), the laptop mostly head to head (~3x), evaluation bursts on the main machine (~1,400 games an hour, paired, one window) against a control that goes on from the same checkpoint for the same budget |
| Upgrades in the shaping | ours (2026-11-12) | --research-value 1 (fgself-213): more upgrades, -11 against insane; not kept |
| Every training game in the real game | ours (bold changes) | --map "(2)EchoIsles" (fgself-212): read as level against a high-point start, then best against insane in a paired evaluation burst (0.86, +8 over fgself-201/142.7M): the base since 2026-11-13 |
| Half the AI games in the real game | ours (bold changes, 2026-11-06) | --real-share 0.5 (fgself-201): +11 points against insane over the continuation, kept as the base |
| The strongest checkpoint as a league member | the user (2026-11-11) | fgself-197's peak fixed in the league since fgself-201's resume |
| The heroes' spells taught | the user (2026-11-11) | --spell-lesson, with a mana cost or without (fgself-208 to -210): the heroes cast ~25-40 spells a game, -10 to -15 against insane, not kept |
| Make sure the agent can do everything | the user (2026-11-06) | --full-vocab 1: every order a melee player can give, 13 exclusions with reasons (order coverage); fgself-196 kept, level |
| What the agent can't observe | the user (2026-11-06) | --obs-full 1: buffs, corpses, items, area spells, enemy upgrades (observation coverage); fgself-197 kept, wins past 18 minutes 16% → 46% |
| BF16 inference | the user (2026-11-06) | --infer-precision bf16-mixed and finer inference rows, part of ~4,300 → ~5,800 steps/s |
| Bold changes over one-flag tweaks | the user (2026-11-06, 2026-11-12) | league-first, dry mains, the full vocabulary, obs-full, real-game training |
| AlphaStar's league first, a league with history | AlphaStar | the AI in 40% of the games (fgself-194): -13 against insane; snapshots spanning hours (fgself-199): level |
| PFSP with data | ours (a bug finder, 2026-11-06) | --pfsp-data (fgself-198): level |
| Dry mains to force expansions | ours (bold changes) | the dry scenario (fgself-195): no expansions in evaluation |
| Expansion lessons with the place taught | a bug finder (2026-11-10) | --expand-lesson 0.5 (fgself-202): pulled hard toward early expansions, 40% against insane |
| More maps in the training pool | the user (2026-11-01) | fgself-178: level off Echo Isles, worse on it; not kept |
| The value sees the tax | the user (2026-11-01) | --value-tax 1 (fgself-184): 41% head to head, off |
| An EMA anchor | the user (2026-11-01) | --ref-ema (fgself-191, -192): level, no wider |
| Shop items in the targets | the user (2026-11-02) | --z-shop 3 (fgself-183): 3 items a game bought, -22 against insane |
| Training games on a second machine | the user (2026-10-24) | the games box: 56, now 64 training games; ~4,450 steps/s with it (fgself-168) |
| Narrow targets, more kinds of conditioning | the user | fgself-109: level; the easy kinds followed, the hard ones not |
| Rewards by part (GDPO) | the user (LLM RL) | worse (fgself-110, -116) |
| Self-distillation with privileged information (OPSD) | the user (LLM training) | no gain (fgself-112, -121) |
| Neutral shops, the tavern, mercenaries | the user | tavern heroes hired; untaught mercenaries cost real-game wins (--neutral 2); shop items as orders kept (fgself-179: 66% → 77%) |
| A heuristic teacher | the user | the teacher's lessons label the policy's steps (--teach 3), in every run |
| Scenarios, weight averaging | the user (the micro curriculum; model soups) | level (fgself-114; -115, a soup of two runs). The mean of a stretch of one run's checkpoints is in use (above) |
| Skill ratings | the user | a Bradley-Terry rating of every policy: the dashboard's Ratings view |
| A heuristic opponent stronger than the learner | the user (2026-10-25) | fullgame/heuristic.py, in the evaluation games (--eval-heuristic) |
| Diverse play without macro help | the user (2026-10-25) | --macros-off mix,creep,expand (fgself-151): kept, it adapted |
| Targets that limit or forbid units | the user (footman spam) | --z-comp (fgself-154): not kept |
| A lumber lesson | ours (audit 9) | fgself-150: kept |
| Same-page merging of the games' memory | ours | on the games box (KSM) |
| Human data | AlphaStar (Fig. 3E) | a clone of people's play was passive (2026-10-27); people's late-game states are late starts (fgself-168) |
| The target's first hero taught at the altar | ours | --teach 2 (fgself-140): level |
| Strategic variety first, a lower win rate accepted for a while | the user (2026-10-18) | the direction since; targets the policy follows on its own came with the tech-tree builder (fgself-95) |
| Faster games for variety: the duel map, faster rules | the user | echofast training judged on the real game: the best line (fgself-100, 95%) |
| Teach expanding, creeping, picking up items | the user (2026-10-22) | as lessons they failed (fgself-123, -124: armies dragged to the camp, halls built at home); as orders of the agent's own that the policy chooses when (--macros: expand, creep) the first expansions in the real game and the best policy yet (fgself-127 to -129); items off the ground: --macros 2 (fgself-134) |
| A tax on self-play by who is expected to win | the user (as the races' balance, for targets too) | a logistic model of the outcome from both sides' races and targets and the tax; the favoured side taxed by what evens the game out (--win-balance, fgself-136) |
| Search the videos | the user | the dashboard's Videos page and each video's target and heroes beside it (dashboard/videos.py) |
| Audits for silently wrong assumptions | the user ("find issues with our current approach", after the tavern's) | two audit subagents (2026-10-22) found ~14: dead heroes never revived, undead never expanding, the expand order taken by every worker, the newest units cut from the view, a tier refused behind a worker queue, a dead tavern hero lost for good... |
| GRPO from mid-game branches | the user (from LLM RL: DeepSeekMath's group-relative advantages) | level (fgself-96, -97): only ~55% of branches replay exactly (creeps and the AI are not deterministic) |
| Classifier-free guidance toward the target | the user (from image generation) | no use (fgself-106): it followed no better and won 50% instead of 91% |
| AlphaStar's league (self-play, past snapshots, a main exploiter) | AlphaStar | level (fgself-103) |
| Reward decay 0.6^(T / 10 min) | OpenAI Five (App. G, eq. 11) | level (fgself-104) |
| Each sample used once | OpenAI Five (§4.4) | level (fgself-105) |
| More games as the weaker races | ours | undead 79% → 89% (fgself-102) |
| The anchor at the best checkpoint | ours | no slide after the peak (fgself-101) |
| Forward KL on every head, UPGO, a moving anchor, a value that sees the opponent | AlphaStar | forward KL and UPGO: large gains; the opponent's information for the value: kept, level |
| The RTX 3090's BF16 rate unlocked (gpu-unlock), 400 W | the user | matrix multiplies 1.4-1.8× faster; self-play's speed unchanged (CPU-bound) (GPU unlock) |
| More of the game's code faster (Ghidra, REA), wineserver | the user | round 2 of --game-cuts: -18% of a game's main thread, the same simulation |
| The actors' Python in C | the user (the levers of 2026-10-24) | the units' features in C: View.step 1,002 → 556 µs a side-step; with 8 games an actor ~1,665 → ~1,730 steps/s |
| A larger network | ours | 5 layers helped on the duel maps; 8 level, and the user's direction: no more layers without a clear gain |
- AlphaStar, OpenAI Five, SC2LE: see what the papers found.
- GDPO (NVIDIA): rewards normalized one by one before they are summed (explained). SAW: stage-aware dynamic weighting, arXiv 2606.07705.
- GRPO: Z. Shao et al., DeepSeekMath, arXiv 2402.03300.
- Classifier-free guidance: J. Ho and T. Salimans, arXiv 2207.12598.
- Model soups: M. Wortsman et al., arXiv 2203.05482. Task arithmetic: G. Ilharco et al., arXiv 2212.04089.
More maps in the training pool (the user, 2026-11-01: Echo Isles may favour rushing; more maps also open more replays). Straight-line base distances do not single it out (Echo Isles 9,856, Amazonia 9,458, Last Refuge 10,549, Terenas Stand 11,860; walking distances not measured), and people on it play a median 13.8 minutes, so first measure: fgself-168's policy against insane on two or three other stock maps, by win rate and game length, with no training. Needs each game's frame in evaluation (bc.py's game_frame; the run's is set once today). Measured on 2026-11-01: the policy on other maps. Built behind a flag: --map-pool (Several maps).
The value sees the tax (the user, 2026-11-01). Neither network reads a game's tax, so the value cannot tell a taxed AI from an untaxed one. In fgself-173's training games against the AI, untaxed launches (11%) won 17%, taxed ones 43-64%. Within a matchup the tax moves little (standard deviation ~0.07, means 0.10 to 0.45 between matchups), and the value can see the races, so most of it can be learned from the matchup; the untaxed launches are what it cannot tell apart. Between agents 87% of games carry a tax (mean 0.11), and being taxed over 0.2 won 24% to ~33%. If a tax stays, give it to the value only (model.py's priv_proj: a block added to the value head's input, zero at first); never to the policy, since evaluation has no tax. To test as one change, depending on fgself-174. Built behind a flag: --value-tax 1 (two numbers after the privileged block, the opponent's tax and its own, read by a value-only projection, tax_proj, zero at first).
Loosening the KL anchor, after fgself-171 (the user, 2026-11-01: its replays looked more varied; update the anchor, or an EMA anchor?). Without the anchor orc and night elf played more varied, but the macro fell and it lost. Past moves: a rolling anchor (fgself-82, every 200 updates, worse, but without shaping and at doubled income), the anchor at the fork's start (fgself-141, worse and falling). A fixed anchor on a recent checkpoint holds it to one-unit habits, so these come first: (1) --ref-kl 0.01, no code; (2) an EMA anchor (the user's idea: the anchor follows the policy slowly, so it stops sudden jumps without pulling back to fgself-100; risk: it drifts as the macro does, only later); (3) fgself-171's last as a fixed league member (--league-add), which puts its armies in front of the main policy, no code; (4) if the macro is what falls, the anchor on workers and halls only, so army choice is free.