You can deploy this model and start earning money today!

aminwoo/
bughouse-rise-v3

Name: bughouse-rise-v3
Availability: InStock

onnx

chess

bughouse

reinforcement-learning

mcts

alphazero

mit

Views

No views yet

Hivemind — Bughouse RISEv3.3 (cross-board), RL iteration 04

Neural network weights for Hivemind, a UCI engine for Bughouse chess — the four-player variant played on two boards, where captured pieces are passed to your partner. The network drives a Monte Carlo Graph Search (MCGS) that evaluates both boards jointly.

This is RL iteration 04, trained by self-play reinforcement learning.


Architecture	`CrossBoardRiseV3` (RISEv3.3 with cross-board attention)
Parameters	13.57 M (ONNX, BatchNorm folded) · 14.18 M (checkpoint)
Validation loss	1.556
Policy accuracy	82.0 %
Input representation	version 3

Files

File	Size	Purpose
`hivemind-it04-crossboard-risev33-loss1.556-p82.0.onnx`	27.3 MB	Inference — load in ONNX Runtime, or build a TensorRT engine from it
`hivemind-it04-crossboard-risev33-loss1.556-p82.0.tar`	111.4 MB	Training checkpoint — model + optimizer state, for resuming RL

Note on the .tar: despite the extension (inherited from the CrazyAra-derived training code) this is not a tar archive — it is a torch.save zip. Use torch.load, not tar -x.

ONNX signature

Exported with PyTorch 2.9.1 · opset 18 · IR version 8 · dynamic batch dimension.

Input

Name	Shape	Notes
`data`	`[batch, 74, 8, 8]`	74 input planes encoding both boards, pockets, and side-to-move context

Outputs

Name	Shape	Notes
`value`	`[batch, 1]`	Scalar position evaluation
`pi_a`	`[batch, 4672]`	Policy logits, board A
`pi_b`	`[batch, 4672]`	Policy logits, board B
`wdl_out`	`[batch, 3]`	Win / draw / loss head
`moves_left`	`[batch, 1]`	Plies-to-end head

The two policy heads are what make this a bughouse network rather than two independent chess networks: the body reasons over both boards at once and emits a move distribution for each. The 4672 policy size is the standard AlphaZero chess move encoding.

Architecture detail

Body: 15 bottleneck residual blocks at 384 channels (operating width 256, expansion 64). Kernel size 3, widened to 5 at blocks 7, 11, 12, 13. ECA squeeze-excitation at blocks 5, 8, 12, 13, 14.
Cross-board coordination: 2 cross-attention blocks, embedding dim 192, 6 heads — applied after the convolutional body so each board's representation can attend to the other's, including pocket state.
Value head: 16 channels → 512 fully-connected.
Policy heads: 2 shared-weight heads (192 channels), one per board.

Usage

ONNX Runtime

python

import numpy as np, onnxruntime as ort

sess = ort.InferenceSession("hivemind-it04-crossboard-risev33-loss1.556-p82.0.onnx")
planes = np.zeros((1, 74, 8, 8), dtype=np.float32)   # your encoded position
value, pi_a, pi_b, wdl, moves_left = sess.run(None, {"data": planes})

Input planes must be built with the version 3 representation (src/domain/board2planes.py in the engine repo). Feeding a different plane encoding will produce silently wrong evaluations — this is the detail the old -v3.0 filename suffix used to carry.

Resuming training

python

import torch

ck = torch.load("hivemind-it04-crossboard-risev33-loss1.556-p82.0.tar",
                map_location="cpu", weights_only=False)
model.load_state_dict(ck["model_state_dict"])
optimizer.load_state_dict(ck["optimizer_state_dict"])
# ck["training_iteration"] == 13315, ck["evaluation_step"] == 34, ck["batch_steps"] == 400

TensorRT

The engine builds FP16 TensorRT plans from this ONNX at runtime. Prebuilt .engine files are not published here — they are specific to a GPU architecture, TensorRT version, and batch size, so they must be built on the target machine.

Naming

Filenames follow the training loop's convention (hivemind-rl-{iter}-{architecture}-loss{val_loss}-p{policy_acc}):

it04 — RL iteration 4
crossboard-risev33 — architecture
loss1.556 — validation loss
p82.0 — policy accuracy, percent

License

MIT, matching the Hivemind repository.

aminwoo/bughouse-rise-v3