Add SimBa Policy: Simplicity Bias for Scaling Up Parameters in DRL

.github/workflows/ci.yml

@@ -20,7 +20,7 @@ jobs:
     runs-on: ubuntu-latest
     strategy:
       matrix:
-        python-version: ["3.8", "3.9", "3.10", "3.11"]
+        python-version: ["3.9", "3.10", "3.11", "3.12"]
 
     steps:
       - uses: actions/checkout@v3
@@ -52,8 +52,6 @@ jobs:
       - name: Type check
         run: |
           make type
-        # skip mypy, jax doesn't have its latest version for python 3.8
-        if: "!(matrix.python-version == '3.8')"
       - name: Test with pytest
         run: |
           make pytest

README.md

@@ -18,8 +18,11 @@ Implemented algorithms:
 - [Twin Delayed DDPG (TD3)](https://arxiv.org/abs/1802.09477)
 - [Deep Deterministic Policy Gradient (DDPG)](https://arxiv.org/abs/1509.02971)
 - [Batch Normalization in Deep Reinforcement Learning (CrossQ)](https://openreview.net/forum?id=PczQtTsTIX)
+- [Simplicity Bias for Scaling Up Parameters in Deep Reinforcement Learning (SimBa)](https://openreview.net/forum?id=jXLiDKsuDo)
 
 
+Note: parameter resets for off-policy algorithms can be activated by passing a list of timesteps to the model constructor (ex: `param_resets=[int(1e5), int(5e5)]` to reset parameters and optimizers after 100_000 and 500_000 timesteps.
+
 ### Install using pip
 
 For the latest master version:
@@ -132,6 +135,47 @@ Having a higher learning rate for the q-value function is also helpful: `qf_lear
 
 Note: when using the DroQ configuration with CrossQ, you should set `layer_norm=False` as there is already batch normalization.
 
+## Note about SimBa
+
+[SimBa](https://openreview.net/forum?id=jXLiDKsuDo) is a special network architecture for off-policy algorithms (SAC, TQC, ...).
+
+Some recommended hyperparameters (tested on MuJoCo and PyBullet environments):
+```python
+import optax
+
+
+default_hyperparams = dict(
+    n_envs=1,
+    n_timesteps=int(1e6),
+    policy="SimbaPolicy",
+    learning_rate=3e-4,
+    # qf_learning_rate=1e-3,
+    policy_kwargs={
+        "optimizer_class": optax.adamw,
+        # "optimizer_kwargs": {"weight_decay": 0.01},
+        # Note: here [128] represent a residual block, not just a single layer
+        "net_arch": {"pi": [128], "qf": [256, 256]},
+        "n_critics": 2,
+    },
+    learning_starts=10_000,
+    # Important: input normalization using VecNormalize
+    normalize={"norm_obs": True, "norm_reward": False},
+)
+
+hyperparams = {}
+
+# You can also loop gym.registry
+for env_id in [
+    "HalfCheetah-v4",
+    "HalfCheetahBulletEnv-v0",
+    "Ant-v4",
+]:
+    hyperparams[env_id] = default_hyperparams
+```
+
+and then using the RL Zoo script defined above: `python train.py --algo tqc --env HalfCheetah-v4 -c simba.py -P`.
+
+
 ## Benchmark
 
 A partial benchmark can be found on [OpenRL Benchmark](https://wandb.ai/openrlbenchmark/sbx) where you can also find several [reports](https://wandb.ai/openrlbenchmark/sbx/reportlist).

pyproject.toml

@@ -1,8 +1,8 @@
 [tool.ruff]
 # Same as Black.
 line-length = 127
-# Assume Python 3.8
-target-version = "py38"
+# Assume Python 3.9
+target-version = "py39"
 
 [tool.ruff.lint]
 # See https://beta.ruff.rs/docs/rules/
@@ -28,9 +28,7 @@ show_error_codes = true
 
 [tool.pytest.ini_options]
 # Deterministic ordering for tests; useful for pytest-xdist.
-env = [
-	"PYTHONHASHSEED=0"
-]
+env = ["PYTHONHASHSEED=0"]
 
 filterwarnings = [
     # Tensorboard warnings
@@ -41,7 +39,7 @@ filterwarnings = [
     "ignore:rich is experimental",
 ]
 markers = [
-    "expensive: marks tests as expensive (deselect with '-m \"not expensive\"')"
+    "expensive: marks tests as expensive (deselect with '-m \"not expensive\"')",
 ]
 
 [tool.coverage.run]
@@ -50,4 +48,8 @@ branch = false
 omit = ["tests/*", "setup.py"]
 
 [tool.coverage.report]
-exclude_lines = [ "pragma: no cover", "raise NotImplementedError()", "if typing.TYPE_CHECKING:"]
+exclude_lines = [
+    "pragma: no cover",
+    "raise NotImplementedError()",
+    "if typing.TYPE_CHECKING:",
+]

sbx/__init__.py

@@ -23,11 +23,11 @@ def DroQ(*args, **kwargs):
 
 
 __all__ = [
-    "CrossQ",
     "DDPG",
     "DQN",
     "PPO",
     "SAC",
     "TD3",
     "TQC",
+    "CrossQ",
 ]

sbx/common/jax_layers.py

@@ -1,5 +1,7 @@
-from typing import Any, Callable, Optional, Sequence, Tuple, Union
+from collections.abc import Sequence
+from typing import Any, Callable, Optional, Union
 
+import flax.linen as nn
 import jax
 import jax.numpy as jnp
 from flax.linen.module import Module, compact, merge_param
@@ -8,7 +10,7 @@
 
 PRNGKey = Any
 Array = Any
-Shape = Tuple[int, ...]
+Shape = tuple[int, ...]
 Dtype = Any  # this could be a real type?
 Axes = Union[int, Sequence[int]]
 
@@ -204,3 +206,22 @@ def __call__(self, x, use_running_average: Optional[bool] = None):
             self.bias_init,
             self.scale_init,
         )
+
+
+# Adapted from simba: https://github.com/SonyResearch/simba
+class SimbaResidualBlock(nn.Module):
+    hidden_dim: int
+    activation_fn: Callable[[jnp.ndarray], jnp.ndarray] = nn.relu
+    # "the MLP is structured with an inverted bottleneck, where the hidden
+    # dimension is expanded to 4 *  hidden_dim"
+    scale_factor: int = 4
+    norm_layer: type[nn.Module] = nn.LayerNorm
+
+    @nn.compact
+    def __call__(self, x: jnp.ndarray) -> jnp.ndarray:
+        residual = x
+        x = self.norm_layer()(x)
+        x = nn.Dense(self.hidden_dim * self.scale_factor, kernel_init=nn.initializers.he_normal())(x)
+        x = self.activation_fn(x)
+        x = nn.Dense(self.hidden_dim, kernel_init=nn.initializers.he_normal())(x)
+        return residual + x

sbx/common/off_policy_algorithm.py

@@ -1,6 +1,6 @@
 import io
 import pathlib
-from typing import Any, Dict, List, Optional, Tuple, Type, Union
+from typing import Any, Optional, Union
 
 import jax
 import numpy as np
@@ -17,7 +17,7 @@
 class OffPolicyAlgorithmJax(OffPolicyAlgorithm):
     def __init__(
         self,
-        policy: Type[BasePolicy],
+        policy: type[BasePolicy],
         env: Union[GymEnv, str],
         learning_rate: Union[float, Schedule],
         qf_learning_rate: Optional[float] = None,
@@ -26,13 +26,13 @@ def __init__(
         batch_size: int = 256,
         tau: float = 0.005,
         gamma: float = 0.99,
-        train_freq: Union[int, Tuple[int, str]] = (1, "step"),
+        train_freq: Union[int, tuple[int, str]] = (1, "step"),
         gradient_steps: int = 1,
         action_noise: Optional[ActionNoise] = None,
-        replay_buffer_class: Optional[Type[ReplayBuffer]] = None,
-        replay_buffer_kwargs: Optional[Dict[str, Any]] = None,
+        replay_buffer_class: Optional[type[ReplayBuffer]] = None,
+        replay_buffer_kwargs: Optional[dict[str, Any]] = None,
         optimize_memory_usage: bool = False,
-        policy_kwargs: Optional[Dict[str, Any]] = None,
+        policy_kwargs: Optional[dict[str, Any]] = None,
         tensorboard_log: Optional[str] = None,
         verbose: int = 0,
         device: str = "auto",
@@ -43,7 +43,9 @@ def __init__(
         sde_sample_freq: int = -1,
         use_sde_at_warmup: bool = False,
         sde_support: bool = True,
-        supported_action_spaces: Optional[Tuple[Type[spaces.Space], ...]] = None,
+        stats_window_size: int = 100,
+        param_resets: Optional[list[int]] = None,
+        supported_action_spaces: Optional[tuple[type[spaces.Space], ...]] = None,
     ):
         super().__init__(
             policy=policy,
@@ -62,6 +64,7 @@ def __init__(
             use_sde=use_sde,
             sde_sample_freq=sde_sample_freq,
             use_sde_at_warmup=use_sde_at_warmup,
+            stats_window_size=stats_window_size,
             policy_kwargs=policy_kwargs,
             tensorboard_log=tensorboard_log,
             verbose=verbose,
@@ -74,11 +77,25 @@ def __init__(
         self.key = jax.random.PRNGKey(0)
         # Note: we do not allow schedule for it
         self.qf_learning_rate = qf_learning_rate
+        self.param_resets = param_resets
+        self.reset_idx = 0
+
+    def _maybe_reset_params(self) -> None:
+        # Maybe reset the parameters
+        if (
+            self.param_resets
+            and self.reset_idx < len(self.param_resets)
+            and self.num_timesteps >= self.param_resets[self.reset_idx]
+        ):
+            # Note: we are not resetting the entropy coeff
+            assert isinstance(self.qf_learning_rate, float)
+            self.key = self.policy.build(self.key, self.lr_schedule, self.qf_learning_rate)
+            self.reset_idx += 1
 
     def _get_torch_save_params(self):
         return [], []
 
-    def _excluded_save_params(self) -> List[str]:
+    def _excluded_save_params(self) -> list[str]:
         excluded = super()._excluded_save_params()
         excluded.remove("policy")
         return excluded

sbx/common/on_policy_algorithm.py

@@ -1,4 +1,4 @@
-from typing import Any, Dict, List, Optional, Tuple, Type, TypeVar, Union
+from typing import Any, Optional, TypeVar, Union
 
 import gymnasium as gym
 import jax
@@ -24,7 +24,7 @@ class OnPolicyAlgorithmJax(OnPolicyAlgorithm):
 
     def __init__(
         self,
-        policy: Union[str, Type[BasePolicy]],
+        policy: Union[str, type[BasePolicy]],
         env: Union[GymEnv, str],
         learning_rate: Union[float, Schedule],
         n_steps: int,
@@ -37,12 +37,12 @@ def __init__(
         sde_sample_freq: int,
         tensorboard_log: Optional[str] = None,
         monitor_wrapper: bool = True,
-        policy_kwargs: Optional[Dict[str, Any]] = None,
+        policy_kwargs: Optional[dict[str, Any]] = None,
         verbose: int = 0,
         seed: Optional[int] = None,
         device: str = "auto",
         _init_setup_model: bool = True,
-        supported_action_spaces: Optional[Tuple[Type[spaces.Space], ...]] = None,
+        supported_action_spaces: Optional[tuple[type[spaces.Space], ...]] = None,
     ):
         super().__init__(
             policy=policy,  # type: ignore[arg-type]
@@ -70,7 +70,7 @@ def __init__(
     def _get_torch_save_params(self):
         return [], []
 
-    def _excluded_save_params(self) -> List[str]:
+    def _excluded_save_params(self) -> list[str]:
         excluded = super()._excluded_save_params()
         excluded.remove("policy")
         return excluded