saferl-lite 0.1.0__py3-none-any.whl
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- agents/__init__.py +0 -0
- agents/constrained_dqn.py +82 -0
- agents/constraints.py +39 -0
- envs/__init__.py +0 -0
- envs/wrappers.py +58 -0
- explainability/__init__.py +0 -0
- explainability/saliency.py +23 -0
- explainability/shap_explainer.py +31 -0
- saferl_lite-0.1.0.dist-info/METADATA +139 -0
- saferl_lite-0.1.0.dist-info/RECORD +13 -0
- saferl_lite-0.1.0.dist-info/WHEEL +5 -0
- saferl_lite-0.1.0.dist-info/licenses/LICENSE +21 -0
- saferl_lite-0.1.0.dist-info/top_level.txt +3 -0
agents/__init__.py
ADDED
File without changes
|
@@ -0,0 +1,82 @@
|
|
1
|
+
# agents/constrained_dqn.py
|
2
|
+
|
3
|
+
import numpy as np
|
4
|
+
import random
|
5
|
+
import torch
|
6
|
+
import torch.nn as nn
|
7
|
+
import torch.optim as optim
|
8
|
+
from collections import deque
|
9
|
+
from agents.constraints import Constraint
|
10
|
+
|
11
|
+
|
12
|
+
class DQNetwork(nn.Module):
|
13
|
+
def __init__(self, input_dim, output_dim):
|
14
|
+
super().__init__()
|
15
|
+
self.net = nn.Sequential(
|
16
|
+
nn.Linear(input_dim, 128),
|
17
|
+
nn.ReLU(),
|
18
|
+
nn.Linear(128, 128),
|
19
|
+
nn.ReLU(),
|
20
|
+
nn.Linear(128, output_dim),
|
21
|
+
)
|
22
|
+
|
23
|
+
def forward(self, x):
|
24
|
+
return self.net(x)
|
25
|
+
|
26
|
+
|
27
|
+
class ConstrainedDQNAgent:
|
28
|
+
def __init__(self, state_dim, action_dim, constraint: Constraint = None):
|
29
|
+
self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
30
|
+
self.q_net = DQNetwork(state_dim, action_dim).to(self.device)
|
31
|
+
self.target_net = DQNetwork(state_dim, action_dim).to(self.device)
|
32
|
+
self.target_net.load_state_dict(self.q_net.state_dict())
|
33
|
+
|
34
|
+
self.optimizer = optim.Adam(self.q_net.parameters(), lr=1e-3)
|
35
|
+
self.memory = deque(maxlen=10000)
|
36
|
+
self.gamma = 0.99
|
37
|
+
self.batch_size = 64
|
38
|
+
|
39
|
+
self.constraint = constraint
|
40
|
+
self.action_dim = action_dim
|
41
|
+
|
42
|
+
def select_action(self, state, epsilon):
|
43
|
+
if random.random() < epsilon:
|
44
|
+
return random.randint(0, self.action_dim - 1)
|
45
|
+
state = torch.tensor(state, dtype=torch.float32).to(self.device)
|
46
|
+
with torch.no_grad():
|
47
|
+
q_values = self.q_net(state)
|
48
|
+
return q_values.argmax().item()
|
49
|
+
|
50
|
+
def store_transition(self, s, a, r, s_next, done):
|
51
|
+
self.memory.append((s, a, r, s_next, done))
|
52
|
+
|
53
|
+
def update(self):
|
54
|
+
if len(self.memory) < self.batch_size:
|
55
|
+
return
|
56
|
+
batch = random.sample(self.memory, self.batch_size)
|
57
|
+
s, a, r, s_next, done = zip(*batch)
|
58
|
+
|
59
|
+
s = torch.tensor(s, dtype=torch.float32).to(self.device)
|
60
|
+
a = torch.tensor(a).to(self.device)
|
61
|
+
r = torch.tensor(r, dtype=torch.float32).to(self.device)
|
62
|
+
s_next = torch.tensor(s_next, dtype=torch.float32).to(self.device)
|
63
|
+
done = torch.tensor(done, dtype=torch.float32).to(self.device)
|
64
|
+
|
65
|
+
q_values = self.q_net(s).gather(1, a.unsqueeze(1)).squeeze()
|
66
|
+
with torch.no_grad():
|
67
|
+
target_q = r + self.gamma * self.target_net(s_next).max(1)[0] * (1 - done)
|
68
|
+
|
69
|
+
loss = nn.functional.mse_loss(q_values, target_q)
|
70
|
+
self.optimizer.zero_grad()
|
71
|
+
loss.backward()
|
72
|
+
self.optimizer.step()
|
73
|
+
|
74
|
+
def apply_constraint(self, state, action, reward):
|
75
|
+
if self.constraint:
|
76
|
+
penalty = self.constraint.compute_penalty(state, action, reward)
|
77
|
+
return reward - penalty, penalty
|
78
|
+
return reward, 0.0
|
79
|
+
|
80
|
+
def reset_constraints(self):
|
81
|
+
if self.constraint:
|
82
|
+
self.constraint.reset()
|
agents/constraints.py
ADDED
@@ -0,0 +1,39 @@
|
|
1
|
+
# agents/constraints.py
|
2
|
+
|
3
|
+
from abc import ABC, abstractmethod
|
4
|
+
|
5
|
+
|
6
|
+
class Constraint(ABC):
|
7
|
+
"""Abstract base class for constraints in SafeRL agents."""
|
8
|
+
|
9
|
+
@abstractmethod
|
10
|
+
def compute_penalty(self, state, action, reward) -> float:
|
11
|
+
"""Return penalty value for a given step."""
|
12
|
+
pass
|
13
|
+
|
14
|
+
def reset(self):
|
15
|
+
"""Optional: reset internal counters for new episode."""
|
16
|
+
pass
|
17
|
+
|
18
|
+
|
19
|
+
class ActionBudgetConstraint(Constraint):
|
20
|
+
def __init__(self, max_actions: int):
|
21
|
+
self.max_actions = max_actions
|
22
|
+
self.counter = 0
|
23
|
+
|
24
|
+
def compute_penalty(self, state, action, reward) -> float:
|
25
|
+
self.counter += 1
|
26
|
+
return 1.0 if self.counter > self.max_actions else 0.0
|
27
|
+
|
28
|
+
def reset(self):
|
29
|
+
self.counter = 0
|
30
|
+
|
31
|
+
|
32
|
+
class EnergyPenaltyConstraint(Constraint):
|
33
|
+
def __init__(self, energy_fn, max_energy):
|
34
|
+
self.energy_fn = energy_fn
|
35
|
+
self.max_energy = max_energy
|
36
|
+
|
37
|
+
def compute_penalty(self, state, action, reward) -> float:
|
38
|
+
energy = self.energy_fn(state, action)
|
39
|
+
return max(0.0, energy - self.max_energy)
|
envs/__init__.py
ADDED
File without changes
|
envs/wrappers.py
ADDED
@@ -0,0 +1,58 @@
|
|
1
|
+
# envs/wrappers.py
|
2
|
+
|
3
|
+
import gymnasium as gym
|
4
|
+
import numpy as np
|
5
|
+
|
6
|
+
|
7
|
+
class SafeEnvWrapper(gym.Wrapper):
|
8
|
+
def __init__(self, env, max_force: float = None, max_energy: float = None):
|
9
|
+
super().__init__(env)
|
10
|
+
self.max_force = max_force
|
11
|
+
self.max_energy = max_energy
|
12
|
+
self.violation_log = []
|
13
|
+
self.episode_log = []
|
14
|
+
|
15
|
+
def reset(self, **kwargs):
|
16
|
+
obs, info = self.env.reset(**kwargs)
|
17
|
+
self.violation_log.clear()
|
18
|
+
self.episode_log.clear()
|
19
|
+
return obs, info
|
20
|
+
|
21
|
+
def step(self, action):
|
22
|
+
obs, reward, terminated, truncated, info = self.env.step(action)
|
23
|
+
done = terminated or truncated
|
24
|
+
|
25
|
+
violation = 0.0
|
26
|
+
|
27
|
+
# Constraint: limit max force (CartPole)
|
28
|
+
if self.max_force is not None:
|
29
|
+
force = self._get_force(action)
|
30
|
+
if abs(force) > self.max_force:
|
31
|
+
violation = 1.0
|
32
|
+
reward -= 1.0 # penalize
|
33
|
+
|
34
|
+
# Note: remove _get_energy for now (MountainCar support will come later)
|
35
|
+
|
36
|
+
self.violation_log.append(violation)
|
37
|
+
self.episode_log.append(
|
38
|
+
{
|
39
|
+
"obs": obs,
|
40
|
+
"action": action,
|
41
|
+
"reward": reward,
|
42
|
+
"violation": violation,
|
43
|
+
}
|
44
|
+
)
|
45
|
+
|
46
|
+
return obs, reward, terminated, truncated, info
|
47
|
+
|
48
|
+
def _get_force(self, action):
|
49
|
+
# Assumes CartPole: force is ยฑ10
|
50
|
+
return 10.0 if action == 1 else -10.0
|
51
|
+
|
52
|
+
def _get_energy(self, prev_obs, action, next_obs):
|
53
|
+
if prev_obs is None:
|
54
|
+
return 0.0
|
55
|
+
# Simplified energy calculation: KE + PE
|
56
|
+
velocity = next_obs[1]
|
57
|
+
height = np.cos(3 * next_obs[0]) # approximates potential
|
58
|
+
return 0.5 * velocity**2 + 9.8 * height
|
File without changes
|
@@ -0,0 +1,23 @@
|
|
1
|
+
# explainability/saliency.py
|
2
|
+
|
3
|
+
import torch
|
4
|
+
from captum.attr import Saliency
|
5
|
+
|
6
|
+
|
7
|
+
class SaliencyExplainer:
|
8
|
+
def __init__(self, model, device="cpu"):
|
9
|
+
self.model = model
|
10
|
+
self.device = device
|
11
|
+
self.saliency = Saliency(self.model)
|
12
|
+
|
13
|
+
def explain(self, state_tensor, target_action: int):
|
14
|
+
"""
|
15
|
+
state_tensor: 1D torch tensor (state) on correct device
|
16
|
+
target_action: int, index of the action you want to explain
|
17
|
+
Returns: 1D saliency values (array) for input features
|
18
|
+
"""
|
19
|
+
state_tensor = state_tensor.unsqueeze(
|
20
|
+
0
|
21
|
+
).requires_grad_() # Shape: [1, input_dim]
|
22
|
+
attr = self.saliency.attribute(state_tensor, target=target_action)
|
23
|
+
return attr.squeeze().detach().cpu().numpy()
|
@@ -0,0 +1,31 @@
|
|
1
|
+
# explainability/shap_explainer.py
|
2
|
+
|
3
|
+
import shap
|
4
|
+
import numpy as np
|
5
|
+
import torch
|
6
|
+
|
7
|
+
|
8
|
+
class SHAPExplainer:
|
9
|
+
def __init__(self, model, input_dim, device="cpu"):
|
10
|
+
"""
|
11
|
+
model: a function that maps np.array -> Q-values
|
12
|
+
input_dim: size of observation space
|
13
|
+
"""
|
14
|
+
self.input_dim = input_dim
|
15
|
+
self.device = device
|
16
|
+
|
17
|
+
def model_wrapper(x_np):
|
18
|
+
x_tensor = torch.tensor(x_np, dtype=torch.float32).to(self.device)
|
19
|
+
with torch.no_grad():
|
20
|
+
return model(x_tensor).cpu().numpy()
|
21
|
+
|
22
|
+
self.explainer = shap.Explainer(
|
23
|
+
model_wrapper, shap.maskers.Independent(np.zeros((1, input_dim)))
|
24
|
+
)
|
25
|
+
|
26
|
+
def explain(self, state):
|
27
|
+
"""
|
28
|
+
state: np.array of shape (input_dim,)
|
29
|
+
Returns: SHAP values for each input dimension
|
30
|
+
"""
|
31
|
+
return self.explainer(np.array([state]))
|
@@ -0,0 +1,139 @@
|
|
1
|
+
Metadata-Version: 2.4
|
2
|
+
Name: saferl-lite
|
3
|
+
Version: 0.1.0
|
4
|
+
Summary: A lightweight, explainable, and constrained reinforcement learning toolkit.
|
5
|
+
Home-page: https://github.com/satyamcser/saferl-lite
|
6
|
+
Author: Satyam Mishra
|
7
|
+
Author-email: satyam@example.com
|
8
|
+
Project-URL: Documentation, https://satyamcser.github.io/saferl-lite/
|
9
|
+
Project-URL: Source, https://github.com/satyamcser/saferl-lite
|
10
|
+
Project-URL: Bug Tracker, https://github.com/satyamcser/saferl-lite/issues
|
11
|
+
Classifier: Programming Language :: Python :: 3
|
12
|
+
Classifier: License :: OSI Approved :: MIT License
|
13
|
+
Classifier: Operating System :: OS Independent
|
14
|
+
Classifier: Intended Audience :: Science/Research
|
15
|
+
Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
|
16
|
+
Requires-Python: >=3.8
|
17
|
+
Description-Content-Type: text/markdown
|
18
|
+
License-File: LICENSE
|
19
|
+
Requires-Dist: gym
|
20
|
+
Requires-Dist: gymnasium
|
21
|
+
Requires-Dist: numpy
|
22
|
+
Requires-Dist: torch
|
23
|
+
Requires-Dist: matplotlib
|
24
|
+
Requires-Dist: seaborn
|
25
|
+
Requires-Dist: pre-commit
|
26
|
+
Requires-Dist: flake8
|
27
|
+
Requires-Dist: pyyaml
|
28
|
+
Requires-Dist: shap
|
29
|
+
Requires-Dist: captum
|
30
|
+
Requires-Dist: typer
|
31
|
+
Requires-Dist: scikit-learn
|
32
|
+
Requires-Dist: pandas
|
33
|
+
Requires-Dist: pytest
|
34
|
+
Requires-Dist: pytest-cov
|
35
|
+
Requires-Dist: coverage
|
36
|
+
Requires-Dist: mkdocs
|
37
|
+
Requires-Dist: wandb
|
38
|
+
Requires-Dist: mkdocs>=1.5
|
39
|
+
Requires-Dist: mkdocs-material>=9.5
|
40
|
+
Requires-Dist: mkdocstrings[python]
|
41
|
+
Dynamic: author
|
42
|
+
Dynamic: author-email
|
43
|
+
Dynamic: classifier
|
44
|
+
Dynamic: description
|
45
|
+
Dynamic: description-content-type
|
46
|
+
Dynamic: home-page
|
47
|
+
Dynamic: license-file
|
48
|
+
Dynamic: project-url
|
49
|
+
Dynamic: requires-dist
|
50
|
+
Dynamic: requires-python
|
51
|
+
Dynamic: summary
|
52
|
+
|
53
|
+
# ๐ SafeRL-Lite
|
54
|
+
|
55
|
+
A **lightweight, explainable, and modular** Python library for **Constrained Reinforcement Learning (Safe RL)** with real-time **SHAP & saliency-based explainability**, custom metrics, and Gym-compatible wrappers.
|
56
|
+
|
57
|
+
<p align="center">
|
58
|
+
<img src="https://img.shields.io/github/license/satyamcser/saferl-lite?style=flat-square">
|
59
|
+
<img src="https://img.shields.io/github/stars/satyamcser/saferl-lite?style=flat-square">
|
60
|
+
<img src="https://img.shields.io/pypi/v/saferl-lite?style=flat-square">
|
61
|
+
<img src="https://img.shields.io/github/actions/workflow/status/satyamcser/saferl-lite/ci.yml?branch=main&style=flat-square">
|
62
|
+
</p>
|
63
|
+
|
64
|
+
---
|
65
|
+
|
66
|
+
## ๐ Overview
|
67
|
+
|
68
|
+
**SafeRL-Lite** empowers reinforcement learning agents to act under **safety constraints**, while remaining **interpretable** and **modular** for fast experimentation. It wraps standard Gym environments and DQN-based agents with:
|
69
|
+
|
70
|
+
- โ
Safety constraint logic
|
71
|
+
- ๐ Visual explainability (SHAP, saliency maps)
|
72
|
+
- ๐ Violation and reward tracking
|
73
|
+
- ๐งช Built-in testing and evaluations
|
74
|
+
|
75
|
+
---
|
76
|
+
|
77
|
+
## ๐ง Installation
|
78
|
+
|
79
|
+
> ๐ฆ PyPI (coming soon)
|
80
|
+
```bash
|
81
|
+
pip install saferl-lite
|
82
|
+
```
|
83
|
+
|
84
|
+
## ๐ ๏ธ From source:
|
85
|
+
|
86
|
+
```bash
|
87
|
+
git clone https://github.com/satyamcser/saferl-lite.git
|
88
|
+
cd saferl-lite
|
89
|
+
pip install -e .
|
90
|
+
```
|
91
|
+
|
92
|
+
## ๐ Quickstart
|
93
|
+
Train a constrained DQN agent with saliency-based explainability:
|
94
|
+
|
95
|
+
```bash
|
96
|
+
python train.py --env CartPole-v1 --constraint pole_angle --explain shap
|
97
|
+
```
|
98
|
+
|
99
|
+
๐น This:
|
100
|
+
|
101
|
+
- Adds a pole-angle constraint wrapper to the Gym env
|
102
|
+
|
103
|
+
- Logs violations
|
104
|
+
|
105
|
+
- Displays SHAP or saliency explanations for agent decisions
|
106
|
+
|
107
|
+
## ๐ง Features
|
108
|
+
#### โ
Constrained RL
|
109
|
+
- Add custom constraints via wrapper or logic class
|
110
|
+
|
111
|
+
- Violation logging and reward shaping
|
112
|
+
|
113
|
+
- Safe vs unsafe episode tracking
|
114
|
+
|
115
|
+
#### ๐ Explainability
|
116
|
+
- SaliencyExplainer โ gradient-based visual heatmaps
|
117
|
+
|
118
|
+
- SHAPExplainer โ feature contribution values per decision
|
119
|
+
|
120
|
+
- Compatible with any PyTorch-based agent
|
121
|
+
|
122
|
+
#### ๐ Metrics
|
123
|
+
- Constraint violation rate
|
124
|
+
|
125
|
+
- Episode reward
|
126
|
+
|
127
|
+
- Cumulative safe reward
|
128
|
+
|
129
|
+
- Action entropy & temporal behavior stats
|
130
|
+
|
131
|
+
#### ๐ Modularity
|
132
|
+
- Swap out agents, constraints, evaluators, or explainers
|
133
|
+
|
134
|
+
- Supports Gym environments
|
135
|
+
|
136
|
+
- Configurable training pipeline
|
137
|
+
|
138
|
+
## ๐ Citation
|
139
|
+
Coming soon after arXiv/preprint release.
|
@@ -0,0 +1,13 @@
|
|
1
|
+
agents/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
|
2
|
+
agents/constrained_dqn.py,sha256=dkVfgxBUEpT1gh4L2PJBXvwqGbIGFj8mYPFnacqNBgU,2814
|
3
|
+
agents/constraints.py,sha256=en8uB2gluDI6JDsz96lM0yUnFA-0DB9m_a3ycrVky8c,1075
|
4
|
+
envs/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
|
5
|
+
envs/wrappers.py,sha256=rfk3cfsTsfD8NqUjEcJ-o7XGMmkBBHt5kfaCiE3AgAw,1749
|
6
|
+
explainability/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
|
7
|
+
explainability/saliency.py,sha256=EpvrpkRZWqYqd3lkRIkfIbJ0pw7G_hJ8GEiVfgPo88U,767
|
8
|
+
explainability/shap_explainer.py,sha256=Tj-fP947z8ixFdWRXHdR6D3a_wtznGN5x-DomU34xbc,883
|
9
|
+
saferl_lite-0.1.0.dist-info/licenses/LICENSE,sha256=WRhQPkdFDzbMFEhvoaq9gSNnbsy0lhSC8tFH3stLntY,1070
|
10
|
+
saferl_lite-0.1.0.dist-info/METADATA,sha256=k9EwE0Clqv-yIANmGdhJPemW4EhBI9kqAnw6xc74WJE,3868
|
11
|
+
saferl_lite-0.1.0.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
|
12
|
+
saferl_lite-0.1.0.dist-info/top_level.txt,sha256=f1IuezLA5sRnSuKZbl-VrS_Hh9pekOW2smLrpJLuiGg,27
|
13
|
+
saferl_lite-0.1.0.dist-info/RECORD,,
|
@@ -0,0 +1,21 @@
|
|
1
|
+
MIT License
|
2
|
+
|
3
|
+
Copyright (c) 2025 Satyam Mishra
|
4
|
+
|
5
|
+
Permission is hereby granted, free of charge, to any person obtaining a copy
|
6
|
+
of this software and associated documentation files (the "Software"), to deal
|
7
|
+
in the Software without restriction, including without limitation the rights
|
8
|
+
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
|
9
|
+
copies of the Software, and to permit persons to whom the Software is
|
10
|
+
furnished to do so, subject to the following conditions:
|
11
|
+
|
12
|
+
The above copyright notice and this permission notice shall be included in all
|
13
|
+
copies or substantial portions of the Software.
|
14
|
+
|
15
|
+
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
|
16
|
+
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
|
17
|
+
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
|
18
|
+
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
|
19
|
+
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
|
20
|
+
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
|
21
|
+
SOFTWARE.
|