The General Intelligent AgeNt Trainer (GIANT) is an open-source project that provides a flexible architecture for training intelligent agents using various machine learning techniques in diverse simulation environments. The platform’s primary objective is to integrate advanced existing solutions, such as metaheuristic frameworks and game engines, to streamline the training and evaluation process.
The current implementation supports training with the Unity game engine and the EARS framework. Unity serves as an evaluation environment where solutions generated by the EARS framework are tested, while EARS functions as a metaheuristic framework that includes several validated optimization algorithms.
GIANT is designed with a highly modular structure, ensuring that each component operates independently and can be replaced or upgraded as needed. This makes the platform valuable for both researchers and game developers, providing tools for efficiently developing and evaluating new solutions.
- Modular and component-oriented architecture
- Support for single- and multi-agent scenarios with coevolution
- Support for heterogeneous individual organizations
- Diverse decision-making systems (BTs, NNs, script-based, manual control, etc.)
- Simulation speedup and multilevel parallelization
- Seven built-in example environments (benchmark)
- Manual agent testing
- Integration with external ML frameworks via API
- Visualization tools and logging
For a detailed description of these functionalities, see GIANT overview docs.
| Version | Release Date | Source (Git branch) | Docs | Download |
|---|---|---|---|---|
| Release 2 | November 11th, 2025 | branch | docs | download |
| Release 1 | April 2nd, 2025 | branch | docs | download |
| Beta | Oktober 8th, 2024 | -- | -- | download |
For older releases, refer to the Releases page.
If you use GIANT, please include this citation as a reference to the platform.
@article{smid2026giant,
title = {GIANT: General intelligent AgeNt trainer},
journal = {SoftwareX},
volume = {34},
pages = {102607},
year = {2026},
issn = {2352-7110},
doi = {https://doi.org/10.1016/j.softx.2026.102607},
url = {https://www.sciencedirect.com/science/article/pii/S2352711026001007},
author = {Marko Šmid and Miha Ravber},
keywords = {Machine learning, Evolutionary algorithms, Genetic programming, Optimization, Games, Multi-agent systems}
}
- Marko Šmid, Matej Moravec, and Miha Ravber. Msp and ncfp: Novel bloat control methods for genetic programming. IEEE Slovenia section, 2025.