gym-saturation 0.10.1

Gymnasium environments for saturation provers

gym-saturation

gym-saturation is a collection of Gymnasium environments for reinforcement learning (RL) agents striving to prove theorems. Currently, only theorems written in TPTP library formal language are supported.

There are two environments in gym-saturation following the same API: SaturationEnv: VampireEnv is a wrapper around a recent Vampire prover, and IProverEnv relies on a stable version of iProver.

In contrast to monolithic architecture of a typical Automated Theorem Prover (ATP), gym-saturation gives different agents opportunities to select clauses themselves and train from their experience. Combined with a particular agent, gym-saturation can work as an ATP.

gym-saturation can be interesting for RL practitioners willing to apply their experience to theorem proving without coding all the logic-related stuff themselves. It also can be useful for automated deduction researchers who want to create an RL-empowered ATP.

How to Install

Attention!

If you want to use VampireEnv you should have a Vampire binary on your machine. For example, download the latest release.

To use IProverEnv, please download a stable iProver release or build it from this commit.

The best way to install this package is to use pip:

pip install gym-saturation

Another option is to use conda:

conda install -c conda-forge gym-saturation

One can also run it in a Docker container (pre-packed with vampire and iproveropt binaries):

docker build -t gym-saturation https://github.com/inpefess/gym-saturation.git
docker run -it --rm -p 8888:8888 gym-saturation jupyter-lab --ip=0.0.0.0 --port=8888

How to use

One can use gym-saturation environments as any other Gymnasium environment:

import gym_saturation
import gymnasium

env = gymnasium.make("Vampire-v0")  # or "iProver-v0"
# skip this line to use the default problem
env.set_task("a-TPTP-problem-filename")
observation, info = env.reset()
terminated, truncated = False, False
while not (terminated or truncated):
    # apply policy (a valid random action here)
    action = env.action_space.sample(mask=observation["action_mask"])
    observation, reward, terminated, truncated, info = env.step(action)
env.close()

Or have a look at the basic tutorial.

For a bit more comprehensive experiments, please navigate the documentation page.

How to Contribute

Please follow the contribution guide while adhering to the code of conduct.

More documentation

More documentation can be found here.