5.0c Env - Guerrilla Checkers

config/guerrillacheckers.ini

@@ -0,0 +1,53 @@
+[base]
+env_name = guerrillacheckers
+# One epoch is 4096 agents * 64 horizon = 262,144 agent steps. Save every
+# 20 epochs so the historical pool receives a new policy about every 5.24M.
+checkpoint_interval = 20
+
+[selfplay]
+enabled = 1
+max_size = 100
+seed = 42
+# Refresh each frozen bank about as often as snapshots are produced. Loading is
+# deferred until all tagged games reach an episode boundary.
+opp_timeout_steps = 5_000_000
+
+[vec]
+total_agents = 4096
+num_buffers = 8
+num_threads = 8
+num_frozen_banks = 1
+# 80% current-v-current envs and 20% current-v-historical envs.
+frozen_bank_pct = 0.2
+action_mask_size = 256
+
+[env]
+# hard step cap; on timeout the Guerrilla is deemed to have failed (COIN wins).
+max_episode_length = 256
+render_fps = 12
+# selfplay = 1: two policy slots alternate sides. In historical-pool envs,
+# slot 0 is the primary and slot 1 is the frozen opponent.
+# selfplay = 0: policy plays `side` against a built-in `opponent` bot.
+selfplay = 1
+# Side played by slot 0 (the primary policy) in selfplay, or by the agent in
+# bot mode: 0 = random per episode, 1 = Guerrilla, 2 = COIN.
+side = 0
+# opponent bot when selfplay = 0: 0 = random, 1 = greedy (capture-maximizing), 2 = mcts.
+opponent = 1
+# MCTS opponent knobs (used when opponent = 2). Higher iterations = stronger, slower.
+mcts_iterations = 256
+mcts_exploration = 0.7
+# rollout policy: 0 = random (faithful UCT), 1 = greedy (stronger, esp. as guerrilla).
+mcts_rollout = 1
+
+[policy]
+hidden_size = 128
+num_layers = 2
+expansion_factor = 1
+
+[train]
+total_timesteps = 100_000_000
+gamma = 0.98
+# This env has a sparse 256-way action head; action masks handle legality, and
+# low entropy pressure keeps exploration from fighting the mask too hard.
+ent_coef = 0.0001

ocean/guerrillacheckers/README.md

@@ -0,0 +1,59 @@
+# Guerrilla Checkers
+
+The standalone client and tournament preserve the original Puffer 40 as the
+baseline for new standard 5c self-play models.
+
+| Checkpoint | Purpose | SHA-256 |
+| --- | --- | --- |
+| `guerrillacheckers_weights.bin` | Original Puffer 40 | `f58615069b9a0e50105dd54b4729d366d602fe14b957a0042e4dc88089120398` |
+
+Puffer 40 uses the legacy 587,264-byte biased layout. Newly trained standard
+5c checkpoints use the 586,240-byte bias-free layout with a value-head row.
+The standalone evaluator detects both formats directly.
+
+## Train
+
+The canonical `config/guerrillacheckers.ini` uses genuine alternating self-play
+with randomized sides: slot 0 is the trainable policy, slot 1 is the current or
+historical opponent, and the inactive slot receives one deterministic pass
+action. Both recurrent states observe every game step, following the existing
+Ocean Chess self-play convention. Twenty percent of environments use the
+historical bank. Opponent swaps happen only after all tagged games reach an
+episode boundary.
+
+```sh
+./puffer train guerrillacheckers
+```
+
+## Evaluate
+
+Build the standalone client and compare a native candidate against the original
+Puffer 40 from both sides:
+
+```sh
+./build.sh guerrillacheckers --fast
+./guerrillacheckers --compare-candidate 1000 \
+  .runtime/checkpoints/guerrillacheckers/<run>/<checkpoint>.bin
+```
+
+The recorded baseline field contains 100 games per cell, shown as Guerrilla
+wins - COIN wins.
+
+| Guerrilla / COIN | Random | Greedy | Puffer 40 | MCTS 2K | MCTS 10K |
+| --- | ---: | ---: | ---: | ---: | ---: |
+| Random | 8-92 | 0-100 | 0-100 | 0-100 | 0-100 |
+| Greedy | 98-2 | 33-67 | 1-99 | 0-100 | 0-100 |
+| Puffer 40 | 96-4 | 99-1 | 15-85 | 12-88 | 9-91 |
+| MCTS 2K | 99-1 | 100-0 | 50-50 | 30-70 | 11-89 |
+| MCTS 10K | 100-0 | 100-0 | 84-16 | 81-19 | 65-35 |
+
+The side-specific Bradley-Terry fit is anchored at Elo 1500 across all ten
+side-specific entries:
+
+| Bot | Guerrilla Elo | COIN Elo |
+| --- | ---: | ---: |
+| Random | 338 | 727 |
+| Greedy | 1120 | 1189 |
+| Puffer 40 | 1636 | 1879 |
+| MCTS 2K | 1834 | 1969 |
+| MCTS 10K | 2200 | 2108 |