Getting Started

Prerequisites

Gyozas requires the SCIP solver (version 8+). But latest version should be installed when PySCIPOpt is installed automatically with Gyozas dependencies

Installation

Install gyozas from PyPI:

pip install "gyozas[viz]"

Or install from source for development:

git clone https://github.com/olivierjuan/gyozas.git
cd gyozas
pip install -e .[viz]

Your First Episode

The code below creates a branching environment on random set cover instances and runs a single episode with a trivial "always pick the first candidate" policy.

import gyozas

# 1. Create an instance generator
instances = gyozas.SetCoverGenerator(n_rows=50, n_cols=100, rng=42)

# 2. Build the environment
env = gyozas.Environment(
    instance_generator=instances,
    observation_function=gyozas.NodeBipartite(),
    reward_function=gyozas.NNodes(),
)

# 3. Reset starts a new episode (loads and solves a fresh instance)
obs, action_set, reward, done, info = env.reset()

total_reward = reward
steps = 0

# 4. Step until the solver finishes
while not done:
    action = action_set[0]  # naive policy
    obs, action_set, reward, done, info = env.step(action)
    total_reward += reward
    steps += 1

print(f"Episode finished in {steps} steps, total reward: {total_reward}")

# 5. Always close when done
env.close()

Understanding the Output

Each call to reset() and step() returns five values:

Value	Type	Description
obs	array/tuple	Observation of the current solver state (e.g. bipartite graph features)
action_set	np.ndarray	Valid actions for the next step (variable indices for branching, node IDs for node selection)
reward	float	Reward signal for the transition
done	bool	Whether the episode has ended (solver finished)
info	dict/None	Additional information from the information function

Next Steps

• Read the Concepts page to understand how dynamics, rewards, and observations work together
• Try different reward functions like SolvingTime or LPIterations
• Use a different instance generator like IndependentSetGenerator