Breadway/omniastra
1
0
Fork 0
Code Issues 1 Pull requests Projects Releases Packages Wiki Activity Actions

Support games with up to 6 players (OMNI-26) #27

Merged
Breadway merged 1 commit from feature/six-player-games into main 2026-10-08 19:54:35 +08:00
Conversation 2 Commits 1 Files changed 16 +358 -88
Breadway commented 2026-10-06 20:12:02 +08:00
Owner
Copy link

Closes OMNI-26

Stacked on #25 (base is fix/review-findings, so the diff is just this change). Retarget to main once #25 merges.

What changed

  • MAX_PLAYERS is 6. The value head, masks and per-player resource arrays follow it. The validator accepts 2 to 6 players.
  • One zero-sum payoff rule for any player count (zero_sum_payoffs): a sole winner scores +1 and every other player -1/(n-1). Shared wins, or survivors when only losers are flagged, share the win on the same scale; everyone or nobody winning is a draw. Timeouts and random adjudication use the same rule. Two-player results are unchanged except that both players losing at once is now a draw instead of -1 each.
  • New PRef::Target(i): acting on the player chosen by a Player action target. Before this, a chosen player could not be referenced from an effect at all (the tempo attack_face hits Opp regardless of the target), which made targeted player actions meaningless beyond two players. The validator checks the target exists and is a player target. Opp stays "next seat in turn order" and is documented as such.
  • Evaluation: play_match puts the agent under test in a rotating seat against n-1 copies of the opponent and classifies wins by payoff sign. omnia play and omnia simulate work per seat. Generator measurement reports a win rate per seat (seat_win_rates), scales the side-balance floor by player count, and rotates the seat in the skill checks.
  • games/melee.ron: a six-player free-for-all point race with player-targeted raids, used by tests and as an example.

Tests

Zero-sum payoff rule for 2 to 6 players; the three hand-written games run to completion with 2 to 6 players and zero-sum payoffs; limit of 2 to 6 enforced; player-target raids hit only the chosen seat; validator checks on player-target references; six-player observations and batches through the model; six-player matches (MCTS beats random, random wins about 1 game in 6); six-player generator measurement. The generic engine tests (determinism, replay, relabelling, hidden-information) now also run on melee.

Limits worth knowing

  • No generator family produces more than two players yet; only melee exercises 3+ players in a purpose-built game.
  • Games with deep response chains need limits.max_stack scaled with the player count (gambit with 5 players overflowed the 2-player limit of 8).
  • Win/draw/loss counts in play_match are by payoff sign; the Wilson interval treats a game as a Bernoulli trial, which is only indicative with more than two players.
  • The model input shape changes again (value head width 6), on top of the layout change in #25, so existing checkpoints do not load either way.
Closes OMNI-26 Stacked on #25 (base is `fix/review-findings`, so the diff is just this change). Retarget to `main` once #25 merges. ## What changed - `MAX_PLAYERS` is 6. The value head, masks and per-player resource arrays follow it. The validator accepts 2 to 6 players. - One zero-sum payoff rule for any player count (`zero_sum_payoffs`): a sole winner scores +1 and every other player `-1/(n-1)`. Shared wins, or survivors when only losers are flagged, share the win on the same scale; everyone or nobody winning is a draw. Timeouts and random adjudication use the same rule. Two-player results are unchanged except that both players losing at once is now a draw instead of -1 each. - New `PRef::Target(i)`: acting on the player chosen by a `Player` action target. Before this, a chosen player could not be referenced from an effect at all (the tempo `attack_face` hits `Opp` regardless of the target), which made targeted player actions meaningless beyond two players. The validator checks the target exists and is a player target. `Opp` stays "next seat in turn order" and is documented as such. - Evaluation: `play_match` puts the agent under test in a rotating seat against `n-1` copies of the opponent and classifies wins by payoff sign. `omnia play` and `omnia simulate` work per seat. Generator measurement reports a win rate per seat (`seat_win_rates`), scales the side-balance floor by player count, and rotates the seat in the skill checks. - `games/melee.ron`: a six-player free-for-all point race with player-targeted raids, used by tests and as an example. ## Tests Zero-sum payoff rule for 2 to 6 players; the three hand-written games run to completion with 2 to 6 players and zero-sum payoffs; limit of 2 to 6 enforced; player-target raids hit only the chosen seat; validator checks on player-target references; six-player observations and batches through the model; six-player matches (MCTS beats random, random wins about 1 game in 6); six-player generator measurement. The generic engine tests (determinism, replay, relabelling, hidden-information) now also run on melee. ## Limits worth knowing - No generator family produces more than two players yet; only `melee` exercises 3+ players in a purpose-built game. - Games with deep response chains need `limits.max_stack` scaled with the player count (gambit with 5 players overflowed the 2-player limit of 8). - Win/draw/loss counts in `play_match` are by payoff sign; the Wilson interval treats a game as a Bernoulli trial, which is only indicative with more than two players. - The model input shape changes again (value head width 6), on top of the layout change in #25, so existing checkpoints do not load either way.
- MAX_PLAYERS 4 -> 6 (value head, masks and resource arrays follow)
- one zero-sum payoff rule for any player count (zero_sum_payoffs): sole
  winner +1, others -1/(n-1), shared wins and all-win/all-lose are draws;
  also used for timeouts and random adjudication
- PRef::Target(i): act on the player chosen by an action target (Opp is only
  the next seat), with validator checks
- play_match, omnia play/simulate and generator measurement work per seat for
  n players (rotating seat, win = positive payoff, per-seat win rates, side
  balance floor scaled by player count)
- games/melee.ron: six-player free-for-all example, plus tests for 2 to 6
  players across engine, validator, model batching, matches and measurement
Breadway commented 2026-10-08 19:53:22 +08:00
Author
Owner
Copy link

Review (board cleanup 2026-10-08)

Verified on top of #25: the build is clean and all workspace tests pass, including the new six-player engine, match, measurement and model tests.

  • zero_sum_payoffs gives the same results as the old code for 2 players: a sole winner gets +1/-1, a shared win or nobody winning is a draw, a sole survivor gets +1, and a timeout tie is a draw. Existing 2p results are unaffected.
  • play_match and measure rotate the agent under test through all seats. The side-balance floor scales with 2/n. The note that scores are only indicative for n>2 is accurate.
  • PRef::Target(i) is checked by the validator against both scope and a Player-target slot. The engine falls back to me on a mismatch, which can only be reached by an invalid def.
  • Format break: MAX_PLAYERS 4->6 widens the value head and every [f32; MAX_PLAYERS] payoff/value array. Older checkpoints and serialized samples will not load, the same as with #25.

Ready to merge after #25. Retarget the base to main first.

## Review (board cleanup 2026-10-08) Verified on top of #25: the build is clean and all workspace tests pass, including the new six-player engine, match, measurement and model tests. - `zero_sum_payoffs` gives the same results as the old code for 2 players: a sole winner gets +1/-1, a shared win or nobody winning is a draw, a sole survivor gets +1, and a timeout tie is a draw. Existing 2p results are unaffected. - `play_match` and `measure` rotate the agent under test through all seats. The side-balance floor scales with 2/n. The note that scores are only indicative for n>2 is accurate. - `PRef::Target(i)` is checked by the validator against both scope and a Player-target slot. The engine falls back to `me` on a mismatch, which can only be reached by an invalid def. - **Format break:** `MAX_PLAYERS` 4->6 widens the value head and every `[f32; MAX_PLAYERS]` payoff/value array. Older checkpoints and serialized samples will not load, the same as with #25. Ready to merge after #25. Retarget the base to main first.
Breadway changed target branch from fix/review-findings to main 2026-10-08 19:54:34 +08:00
Breadway deleted branch feature/six-player-games 2026-10-08 19:54:35 +08:00
Breadway commented 2026-10-08 19:54:52 +08:00
Author
Owner
Copy link

Retargeted from fix/review-findings to main after #25 merged, then merged 2026-10-08. All tests pass, including the six-player ones. MAX_PLAYERS 4->6 widens the value head, so older checkpoints will not load. Closes OMNI-26.

Retargeted from fix/review-findings to main after #25 merged, then merged 2026-10-08. All tests pass, including the six-player ones. MAX_PLAYERS 4->6 widens the value head, so older checkpoints will not load. Closes OMNI-26.
Sign in to join this conversation.
Reviewers
No reviewers
Labels
Clear labels   
bug
No labels
bug
Milestone
Clear milestone
No items
No milestone
Projects
Clear projects
No items
No project
Assignees
Clear assignees
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference
Breadway/omniastra!27
No description provided.