SynergyGrid

The project SYNGrid concerns the design and implementation of a lightweight benchmark environment for evaluating Long-Term Credit Assignment in Artificial Intelligence agents. The benchmark is implemented as a grid-based environment in Python and follows the Gymnasium API standard, allowing reinforcement learning agents to interact with the environment in a consistent and reproducible manner.

---

Requirements

• Python 3.10 – 3.12
• pip

---

Installation

1. Clone the repository

git clone https://github.com/J-manLans/SynergyGrid.git
cd synergygrid

2. Create a virtual environment

python -m venv .venv

Activate it:

• macOS/Linux:

source .venv/bin/activate

• Windows:

.venv\Scripts\activate

3. Install the project

Install in editable mode with development dependencies:

pip install -e .[dev]

---

⚠️ Important Usage Note

Do not run the project without arguments:

python -m synergygrid

This will attempt to evaluate a non-existing model and result in an error.

---

Quick Start

1. Train an agent

python -m synergygrid --timesteps <T> --iterations <I>

This will:

• Train an agent using the default algorithm
• Save model checkpoints during training
• Train the agent for <T> timesteps
  • <T> should be a multiple of 2048, since each training batch contains 2048 steps
• Train the agent for <I> iterations

Example:

python -m synergygrid --timesteps 2048 --iterations 10

---

2. Evaluate a trained agent

python -m synergygrid --run --steps <N>

Replace <N> with the number of steps of a saved model.

Example:

python -m synergygrid --run --steps 20480

---

3. Continue training from a saved model

python -m synergygrid --cont --steps 20480

---

4. Human control mode

python -m synergygrid --human_controls

Play the environment manually using keyboard controls.

---

5. Run without a trained agent (random actions)

python -m synergygrid --run --no-agent

---

Command Line Arguments

Argument	Description
--alg {0,1,2}	Select algorithm index
--run	Run/evaluate a trained agent (default is training)
--no-agent	Use random actions instead of a trained agent
--cont	Continue training from a saved model
--steps <N>	Model checkpoint to load (required with --run or --cont)
--timesteps <N>	Number of timesteps per training iteration
--iterations <N>	Number of training iterations
--human_controls	Enable manual control

---