Metadata-Version: 2.5
Name: survival-rl
Version: 0.0.9
Summary: Explorations into Survival Reinforcement Learning
Project-URL: Homepage, https://pypi.org/project/survival-rl/
Project-URL: Repository, https://github.com/lucidrains/survival-rl
Author-email: Phil Wang <lucidrains@gmail.com>
License: MIT License
        
        Copyright (c) 2026 Phil Wang
        
        Permission is hereby granted, free of charge, to any person obtaining a copy
        of this software and associated documentation files (the "Software"), to deal
        in the Software without restriction, including without limitation the rights
        to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
        copies of the Software, and to permit persons to whom the Software is
        furnished to do so, subject to the following conditions:
        
        The above copyright notice and this permission notice shall be included in all
        copies or substantial portions of the Software.
        
        THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
        IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
        FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
        AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
        LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
        OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
        SOFTWARE.
License-File: LICENSE
Keywords: artificial intelligence,deep learning,reinforcement learning,survival analysis
Classifier: Development Status :: 4 - Beta
Classifier: Intended Audience :: Developers
Classifier: License :: OSI Approved :: MIT License
Classifier: Programming Language :: Python :: 3.10
Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
Requires-Python: >=3.10
Requires-Dist: einops>=0.8.2
Requires-Dist: einx>=0.4.3
Requires-Dist: torch-einops-utils>=0.1.31
Requires-Dist: torch>=2.5
Requires-Dist: x-mlps-pytorch>=0.6.4
Provides-Extra: examples
Provides-Extra: test
Requires-Dist: pytest; extra == 'test'
Description-Content-Type: text/markdown

<img src="./alg1.png" width="350px"></img>

## Survival RL

Explorations into an application of [survival analysis to RL](https://arxiv.org/abs/2605.31273), proposed by Tiofack et al. of Inria earlier this year

## Install

```bash
$ pip install survival-rl
```

## Usage

`HazardCritic` takes a batch of states, actions and events, and predicts for each state-action pair the time bin in which the event is first dwelled upon, conditioned on that event. It is trained by maximum likelihood on the predicted logits, where `reach_event_index` marks the observed time bin, and `-1` (or anything at or past `horizon_cutoff`) is treated as right censored.

```python
import torch
from survival_rl import HazardCritic

critic = HazardCritic(
    dim = 64,
    depth = 2,
    dim_state = 8,
    num_actions = 3,
    pred_time_bins = 32
)

state  = torch.randn(16, 8)
action = torch.randn(16, 3)
event  = torch.randn(16, 8)     # every state and action is conditioned on the event

reach_event_index = torch.randint(0, 32, (16,))
horizon_cutoff    = torch.full((16,), 32)

loss = critic(state, action, event, reach_event_index = reach_event_index, horizon_cutoff = horizon_cutoff)
loss.backward()
```

Once trained, a policy can be extracted for reaching the event, or avoiding it with `negate = True`, given any actor of the form `(state, event) -> action`.

```python
class Actor(torch.nn.Module):
    def __init__(self, dim_state = 8, dim_event = 8, num_actions = 3):
        super().__init__()
        self.net = torch.nn.Sequential(
            torch.nn.Linear(dim_state + dim_event, 64),
            torch.nn.SiLU(),
            torch.nn.Linear(64, num_actions)
        )

    def forward(self, state, event):
        return self.net(torch.cat((state, event), dim = -1))

actor = Actor()

# maximize the value to reach the event

loss = -critic.extract_policy(actor, state, event).mean()
loss.backward()

# or pass `negate = True` to avoid the event instead

loss = -critic.extract_policy(actor, state, event, negate = True).mean()
loss.backward()
```

The reach indices for training are computed from any trajectory data with `compute_first_dwell_time`, which finds the first time each trajectory dwells within `eps` of its event.

```python
import torch
from survival_rl import compute_first_dwell_time

states = torch.randn(4, 64, 8)      # (batch, time, dim)
event  = states[:, -1:]             # (batch, num_queries, dim)

# first dwell of each trajectory at its event - -1 means it was never reached,
# and is right censored at `horizon_cutoff`

reach_event_index, horizon_cutoff = compute_first_dwell_time(event, states = states, eps = 0.25)
```

For a variant where the time bins compete for a single unit of mass through a softmax, use `HazardCriticCompetitive` in place of `HazardCritic`.

## Citations

```bibtex
@inproceedings{Tiofack2026SVL,
    title     = {SVL: Goal-Conditioned Reinforcement Learning as Survival Learning},
    author    = {Franki Nguimatsia Tiofack and Fabian Schramm and Th{\'e}otime Le Hellard and Justin Carpentier},
    booktitle = {International Conference on Machine Learning (ICML)},
    year      = {2026},
    url       = {https://arxiv.org/abs/2604.17551}
}
```

```bibtex
@article{Tiofack2026SRL,
    title   = {Survival Reinforcement Learning: Toward Scalable Self-Supervised RL},
    author  = {Franki Nguimatsia-Tiofack and Fabian Schramm and Th{\'e}otime Le Hellard and Justin Carpentier},
    journal = {arXiv preprint arXiv:2605.31273},
    year    = {2026},
    note    = {19th European Workshop on Reinforcement Learning (EWRL), Oral},
    url     = {https://arxiv.org/abs/2605.31273}
}
```
