File size: 15,351 Bytes
358dd8b | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 | # Categorical DQN (C51)
## Overview
C51 introduces a distributional perspective for DQN: instead of learning a single value for an action, C51 learns to predict a distribution of values for the action. Empirically, C51 demonstrates impressive performance in ALE.
Original papers:
* [A Distributional Perspective on Reinforcement Learning](https://arxiv.org/abs/1707.06887)
## Implemented Variants
| Variants Implemented | Description |
| ----------- | ----------- |
| :material-github: [`c51_atari.py`](https://github.com/vwxyzjn/cleanrl/blob/master/cleanrl/c51_atari.py), :material-file-document: [docs](/rl-algorithms/c51/#c51_ataripy) | For playing Atari games. It uses convolutional layers and common atari-based pre-processing techniques. |
| :material-github: [`c51.py`](https://github.com/vwxyzjn/cleanrl/blob/master/cleanrl/c51.py), :material-file-document: [docs](/rl-algorithms/c51/#c51py) | For classic control tasks like `CartPole-v1`. |
| :material-github: [`c51_atari_jax.py`](https://github.com/vwxyzjn/cleanrl/blob/master/cleanrl/c51_atari_jax.py), :material-file-document: [docs](/rl-algorithms/c51/#c51_atari_jaxpy) | For playing Atari games. It uses convolutional layers and common atari-based pre-processing techniques. |
| :material-github: [`c51_jax.py`](https://github.com/vwxyzjn/cleanrl/blob/master/cleanrl/c51_jax.py), :material-file-document: [docs](/rl-algorithms/c51/#c51_jaxpy) | For classic control tasks like `CartPole-v1`. |
Below are our single-file implementations of C51:
## `c51_atari.py`
The [c51_atari.py](https://github.com/vwxyzjn/cleanrl/blob/master/cleanrl/c51_atari.py) has the following features:
* For playing Atari games. It uses convolutional layers and common atari-based pre-processing techniques.
* Works with the Atari's pixel `Box` observation space of shape `(210, 160, 3)`
* Works with the `Discrete` action space
### Usage
=== "poetry"
```bash
uv pip install ".[atari]"
uv run python cleanrl/c51_atari.py --env-id BreakoutNoFrameskip-v4
uv run python cleanrl/c51_atari.py --env-id PongNoFrameskip-v4
```
=== "pip"
```bash
pip install -r requirements/requirements-atari.txt
python cleanrl/c51_atari.py --env-id BreakoutNoFrameskip-v4
python cleanrl/c51_atari.py --env-id PongNoFrameskip-v4
```
### Explanation of the logged metrics
Running `python cleanrl/c51_atari.py` will automatically record various metrics such as actor or value losses in Tensorboard. Below is the documentation for these metrics:
* `charts/episodic_return`: episodic return of the game
* `charts/SPS`: number of steps per second
* `losses/loss`: the cross entropy loss between the $t$ step state value distribution and the projected $t+1$ step state value distribution
* `losses/q_values`: implemented as `(old_pmfs * q_network.atoms).sum(1)`, which is the sum of the probability of getting returns $x$ (`old_pmfs`) multiplied by $x$ (`q_network.atoms`), averaged over the sample obtained from the replay buffer; useful when gauging if under or over estimation happens
### Implementation details
[c51_atari.py](https://github.com/vwxyzjn/cleanrl/blob/master/cleanrl/c51_atari.py) is based on (Bellemare et al., 2017)[^1] but presents a few implementation differences:
1. (Bellemare et al., 2017)[^1] injects stochaticity by doing "on each frame the environment rejects the agent’s selected action with probability $p = 0.25$", but `c51_atari.py` does not do this
1. `c51_atari.py` use a self-contained evaluation scheme: `c51_atari.py` reports the episodic returns obtained throughout training, whereas (Bellemare et al., 2017)[^1] is trained with `--end-e=0.01` but reported episodic returns using a separate evaluation process with `--end-e=0.001` (See "5.2. State-of-the-Art Results" on page 7).
### Experiment results
To run benchmark experiments, see :material-github: [benchmark/c51.sh](https://github.com/vwxyzjn/cleanrl/blob/master/benchmark/c51.sh). Specifically, execute the following command:
<script src="https://emgithub.com/embed.js?target=https%3A%2F%2Fgithub.com%2Fvwxyzjn%2Fcleanrl%2Fblob%2Fmaster%2Fbenchmark%2Fc51.sh%23L8-L13&style=github&showBorder=on&showLineNumbers=on&showFileMeta=on&showCopy=on"></script>
Below are the average episodic returns for `c51_atari.py`.
| Environment | `c51_atari.py` 10M steps | (Bellemare et al., 2017, Figure 14)[^1] 50M steps | (Hessel et al., 2017, Figure 5)[^3]
| ----------- | ----------- | ----------- | ---- |
| BreakoutNoFrameskip-v4 | 461.86 ± 69.65 | 748 | ~500 at 10M steps, ~600 at 50M steps
| PongNoFrameskip-v4 | 19.46 ± 0.70 | 20.9 | ~20 10M steps, ~20 at 50M steps
| BeamRiderNoFrameskip-v4 | 9592.90 ± 2270.15 | 14,074 | ~12000 10M steps, ~14000 at 50M steps
Note that we save computational time by reducing timesteps from 50M to 10M, but our `c51_atari.py` scores the same or higher than (Mnih et al., 2015)[^1] in 10M steps.
Learning curves:
<div class="grid-container">
<img src="../c51/BeamRiderNoFrameskip-v4.png">
<img src="../c51/BreakoutNoFrameskip-v4.png">
<img src="../c51/PongNoFrameskip-v4.png">
</div>
Tracked experiments and game play videos:
<iframe src="https://wandb.ai/openrlbenchmark/openrlbenchmark/reports/Atari-CleanRL-s-C51--VmlldzoxNzI0NzQ0" style="width:100%; height:500px" title="CleanRL C51 Tracked Experiments"></iframe>
## `c51.py`
The [c51.py](https://github.com/vwxyzjn/cleanrl/blob/master/cleanrl/c51.py) has the following features:
* Works with the `Box` observation space of low-level features
* Works with the `Discrete` action space
* Works with envs like `CartPole-v1`
### Usage
=== "poetry"
```bash
uv run python cleanrl/c51.py --env-id CartPole-v1
```
=== "pip"
```bash
python cleanrl/c51.py --env-id CartPole-v1
```
### Explanation of the logged metrics
See [related docs](/rl-algorithms/c51/#explanation-of-the-logged-metrics) for `c51_atari.py`.
### Implementation details
The [c51.py](https://github.com/vwxyzjn/cleanrl/blob/master/cleanrl/c51.py) shares the same implementation details as [`c51_atari.py`](https://github.com/vwxyzjn/cleanrl/blob/master/cleanrl/c51_atari.py) except the `c51.py` runs with different hyperparameters and neural network architecture. Specifically,
1. `c51.py` uses a simpler neural network as follows:
```python
self.network = nn.Sequential(
nn.Linear(np.array(env.single_observation_space.shape).prod(), 120),
nn.ReLU(),
nn.Linear(120, 84),
nn.ReLU(),
nn.Linear(84, env.single_action_space.n),
)
```
2. `c51.py` runs with different hyperparameters:
```bash
python c51.py --total-timesteps 500000 \
--learning-rate 2.5e-4 \
--buffer-size 10000 \
--gamma 0.99 \
--target-network-frequency 500 \
--max-grad-norm 0.5 \
--batch-size 128 \
--start-e 1 \
--end-e 0.05 \
--exploration-fraction 0.5 \
--learning-starts 10000 \
--train-frequency 10
```
### Experiment results
To run benchmark experiments, see :material-github: [benchmark/c51.sh](https://github.com/vwxyzjn/cleanrl/blob/master/benchmark/c51.sh). Specifically, execute the following command:
<script src="https://emgithub.com/embed.js?target=https%3A%2F%2Fgithub.com%2Fvwxyzjn%2Fcleanrl%2Fblob%2Fmaster%2Fbenchmark%2Fc51.sh%23L2-L6&style=github&showBorder=on&showLineNumbers=on&showFileMeta=on&showCopy=on"></script>
Below are the average episodic returns for `c51.py`.
| Environment | `c51.py` |
| ----------- | ----------- |
| CartPole-v1 | 481.20 ± 20.53 |
| Acrobot-v1 | -87.70 ± 5.52 |
| MountainCar-v0 | -166.38 ± 27.94 |
Note that the C51 has no official benchmark on classic control environments, so we did not include a comparison. That said, our `c51.py` was able to achieve near perfect scores in `CartPole-v1` and `Acrobot-v1`; further, it can obtain successful runs in the sparse environment `MountainCar-v0`.
Learning curves:
<div class="grid-container">
<img src="../c51/CartPole-v1.png">
<img src="../c51/Acrobot-v1.png">
<img src="../c51/MountainCar-v0.png">
</div>
Tracked experiments and game play videos:
<iframe src="https://wandb.ai/openrlbenchmark/openrlbenchmark/reports/Classic-Control-CleanRL-s-C51--VmlldzoxODIwMTE4" style="width:100%; height:500px" title="CleanRL C51 Tracked Experiments"></iframe>
## `c51_atari_jax.py`
The [c51_atari_jax.py](https://github.com/vwxyzjn/cleanrl/blob/master/cleanrl/c51_atari_jax.py) has the following features:
* Uses [Jax](https://github.com/google/jax), [Flax](https://github.com/google/flax), and [Optax](https://github.com/deepmind/optax) instead of `torch`. [c51_atari_jax.py](https://github.com/vwxyzjn/cleanrl/blob/master/cleanrl/c51_atari_jax.py) is roughly 25% faster than [c51_atari.py](https://github.com/vwxyzjn/cleanrl/blob/master/cleanrl/c51_atari.py)
* For playing Atari games. It uses convolutional layers and common atari-based pre-processing techniques.
* Works with the Atari's pixel `Box` observation space of shape `(210, 160, 3)`
* Works with the `Discrete` action space
### Usage
=== "poetry"
```bash
uv pip install ".[atari, jax]"
uv pip install --upgrade "jax[cuda11_cudnn82]==0.4.8" -f https://storage.googleapis.com/jax-releases/jax_cuda_releases.html
uv run python cleanrl/c51_atari_jax.py --env-id BreakoutNoFrameskip-v4
uv run python cleanrl/c51_atari_jax.py --env-id PongNoFrameskip-v4
```
=== "pip"
```bash
pip install -r requirements/requirements-atari.txt
pip install -r requirements/requirements-jax.txt
pip install --upgrade "jax[cuda]==0.3.17" -f https://storage.googleapis.com/jax-releases/jax_cuda_releases.html
python cleanrl/c51_atari_jax.py --env-id BreakoutNoFrameskip-v4
python cleanrl/c51_atari_jax.py --env-id PongNoFrameskip-v4
```
### Explanation of the logged metrics
See [related docs](/rl-algorithms/c51/#explanation-of-the-logged-metrics) for `c51_atari.py`.
### Implementation details
See [related docs](/rl-algorithms/c51/#implementation-details) for `c51_atari.py`.
### Experiment results
To run benchmark experiments, see :material-github: [benchmark/c51.sh](https://github.com/vwxyzjn/cleanrl/blob/master/benchmark/c51.sh). Specifically, execute the following command:
<script src="https://emgithub.com/embed.js?target=https%3A%2F%2Fgithub.com%2Fvwxyzjn%2Fcleanrl%2Fblob%2Fmaster%2Fbenchmark%2Fc51.sh%23L23-L29&style=github&showBorder=on&showLineNumbers=on&showFileMeta=on&showCopy=on"></script>
Below are the average episodic returns for `c51_atari_jax.py`.
| Environment | `c51_atari_jax.py` 10M steps | `c51_atari.py` 10M steps | (Bellemare et al., 2017, Figure 14)[^1] 50M steps | (Hessel et al., 2017, Figure 5)[^3] |
| ----------------------- | ---------------------------- | ------------------------ | ------------------------------------------------- | ------------------------------------- |
| BreakoutNoFrameskip-v4 | 448.56 ± 17.02 | 461.86 ± 69.65 | 748 | ~500 at 10M steps, ~600 at 50M steps |
| PongNoFrameskip-v4 | 19.88 ± 0.31 | 19.46 ± 0.70 | 20.9 | ~20 10M steps, ~20 at 50M steps |
| BeamRiderNoFrameskip-v4 | 9504.91 ± 709.69 | 9592.90 ± 2270.15 | 14,074 | ~12000 10M steps, ~14000 at 50M steps |
Learning curves:
<div class="grid-container">
<img src="../c51/jax/BeamRiderNoFrameskip-v4.png">
<img src="../c51/jax/BeamRiderNoFrameskip-v4-time.png">
<img src="../c51/jax/BreakoutNoFrameskip-v4.png">
<img src="../c51/jax/BreakoutNoFrameskip-v4-time.png">
<img src="../c51/jax/PongNoFrameskip-v4.png">
<img src="../c51/jax/PongNoFrameskip-v4-time.png">
</div>
Tracked experiments and game play videos:
<iframe src="https://wandb.ai/openrlbenchmark/openrlbenchmark/reports/Atari-CleanRL-s-C51-JAX--VmlldzozMjM4MTIy" style="width:100%; height:500px" title="CleanRL C51 Tracked Experiments"></iframe>
## `c51_jax.py`
The [c51_jax.py](https://github.com/vwxyzjn/cleanrl/blob/master/cleanrl/c51_jax.py) has the following features:
* Uses [Jax](https://github.com/google/jax), [Flax](https://github.com/google/flax), and [Optax](https://github.com/deepmind/optax) instead of `torch`. [c51_jax.py](https://github.com/vwxyzjn/cleanrl/blob/master/cleanrl/c51_jax.py) is roughly 55% faster than [c51.py](https://github.com/vwxyzjn/cleanrl/blob/master/cleanrl/c51.py)
* Works with the `Box` observation space of low-level features
* Works with the `Discrete` action space
* Works with envs like `CartPole-v1`
### Usage
=== "poetry"
```bash
uv pip install ".[jax]"
uv pip install --upgrade "jax[cuda11_cudnn82]==0.4.8" -f https://storage.googleapis.com/jax-releases/jax_cuda_releases.html
uv run python cleanrl/c51_jax.py --env-id CartPole-v1
```
=== "pip"
```bash
pip install -r requirements/requirements-jax.txt
pip install --upgrade "jax[cuda]==0.3.17" -f https://storage.googleapis.com/jax-releases/jax_cuda_releases.html
python cleanrl/c51_jax.py --env-id CartPole-v1
```
### Explanation of the logged metrics
See [related docs](/rl-algorithms/c51/#explanation-of-the-logged-metrics) for `c51_atari.py`.
### Implementation details
See [related docs](/rl-algorithms/c51/#implementation-details_1) for `c51.py`.
### Experiment results
To run benchmark experiments, see :material-github: [benchmark/c51.sh](https://github.com/vwxyzjn/cleanrl/blob/master/benchmark/c51.sh). Specifically, execute the following command:
<script src="https://emgithub.com/embed.js?target=https%3A%2F%2Fgithub.com%2Fvwxyzjn%2Fcleanrl%2Fblob%2Fmaster%2Fbenchmark%2Fc51.sh%23L15-L21&style=github&showBorder=on&showLineNumbers=on&showFileMeta=on&showCopy=on"></script>
Below are the average episodic returns for `c51_jax.py`.
| Environment | `c51_jax.py` | `c51.py` |
| -------------- | --------------- | --------------- |
| CartPole-v1 | 491.07 ± 9.70 | 481.20 ± 20.53 |
| Acrobot-v1 | -86.74 ± 2.19 | -87.70 ± 5.52 |
| MountainCar-v0 | -174.30 ± 36.35 | -166.38 ± 27.94 |
Learning curves:
<div class="grid-container">
<img src="../c51/jax/CartPole-v1.png">
<img src="../c51/jax/CartPole-v1-time.png">
<img src="../c51/jax/Acrobot-v1.png">
<img src="../c51/jax/Acrobot-v1-time.png">
<img src="../c51/jax/MountainCar-v0.png">
<img src="../c51/jax/MountainCar-v0-time.png">
</div>
Tracked experiments and game play videos:
<iframe src="https://wandb.ai/openrlbenchmark/openrlbenchmark/reports/Classic-Control-CleanRL-s-C51-JAX--VmlldzozMjM4MTM4" style="width:100%; height:500px" title="CleanRL C51 Tracked Experiments"></iframe>
[^1]:Bellemare, M.G., Dabney, W., & Munos, R. (2017). A Distributional Perspective on Reinforcement Learning. ICML.
[^2]:\[Proposal\] Formal API handling of truncation vs termination. https://github.com/openai/gym/issues/2510
[^3]: Hessel, M., Modayil, J., Hasselt, H.V., Schaul, T., Ostrovski, G., Dabney, W., Horgan, D., Piot, B., Azar, M.G., & Silver, D. (2018). Rainbow: Combining Improvements in Deep Reinforcement Learning. AAAI. |