MachuPikacchu on Nostr: Experimenting with using RL to train an adversarial bitcoin-script generator. It ...
Experimenting with using RL to train an adversarial bitcoin-script generator.
https://github.com/machuPikacchuBTC/bitcoin-rlIt could be useful for testing future opcodes like OP_CAT. It's built on the MuZero architecture with a reward policy derived from Rusty Russell's varops budget calculations.
I don’t have much time to maintain it and for now it’s just for curiosity and learning. No guarantees about accuracy or performance but it can generate valid scripts when trained for about 20-30 minutes on a good CPU. No GPU support for now because the Monte Carlo tree search is sequential and in the current implementation it’s hard to parallelize it so we’re CPU bottlenecked anyway.
Published at
2026-03-28 16:56:12 UTCEvent JSON
{
"id": "ad2658885b0089f065541b3b48951daa2cd24aeb8fd1b8b0a8a1a78dc3667e26",
"pubkey": "1e908fbc1d131c17a87f32069f53f64f45c75f91a2f6d43f8aa6410974da5562",
"created_at": 1774716972,
"kind": 1,
"tags": [
[
"r",
"https://github.com/machuPikacchuBTC/bitcoin-rl"
]
],
"content": "Experimenting with using RL to train an adversarial bitcoin-script generator.\n\nhttps://github.com/machuPikacchuBTC/bitcoin-rl\n\nIt could be useful for testing future opcodes like OP_CAT. It's built on the MuZero architecture with a reward policy derived from Rusty Russell's varops budget calculations.\n\nI don’t have much time to maintain it and for now it’s just for curiosity and learning. No guarantees about accuracy or performance but it can generate valid scripts when trained for about 20-30 minutes on a good CPU. No GPU support for now because the Monte Carlo tree search is sequential and in the current implementation it’s hard to parallelize it so we’re CPU bottlenecked anyway.",
"sig": "83dec34fa609b5a7941d677c04c2f09db924237c59df89ec789013e9ea435c003657991097f0ba05d8764a3c8e53021e6f6174f73743af367ad09273c9ed2860"
}
