pgx
pgx copied to clipboard
๐ฒ Vectorized RL game environments in JAX
|| Non-MinAtar | MinAtar | |:---|:---|:---| | `state.to_svg` | โ | โ | | `pgx.to_svg` | ๐ง | ๐ง | | `pgx.save_svg` | โ | โ (wrapper of `state.save_svg`) |...
- [x] without check(draw) - [x] with check, not continuous(draw) - [x] with continuous check(checked player win)
ๆฅๆฌ่ชใงใใชใใธใฎ่ชฌๆๆใๆธใ - [x] Observation #707 - [x] Action - [x] table - [x] hash table of dds results #192 - [x] example - [x] how to duplicate #771 - [x]...
Is it easy to recover the square info with same operation? - Animal Shogi: `x, sq = jnp.int8(a // 12), jnp.int8(a % 12)` - Chess: `from_, plane = label //...
https://github.com/kenjyoung/MinAtar/commit/da3dc2dc788a8d483e75abc306ebb94d562e2c37
After naive translation, I fixed * `ZeroDivisionError` * `select/cond` to `GameState`: `_remove_stones`, `_merge_chain`, `_apply_action` To make `torch.vmap(step)` work * [ ] fix `_apply_action` * [ ] remove `while_loop`
Or add new Flax baseline and deprecate Haiku baselines