GENESIS v4.0
Generative Evolving Neural Engine for Self-Improving Systems
GENESIS is an experimental multi-agent evolution and learning simulation. The current genesis.py implementation combines neuroevolution, online adaptation, intrinsic-motivation signals, communication, counterfactual evaluation, shared knowledge, hierarchical goals, social prediction, concept formation, and self-narrative mechanisms in one self-contained Python program.
Project status: research prototype. GENESIS explores mechanisms associated with adaptive and self-improving systems; it is not evidence of artificial general intelligence or a formally verified Gödel machine.
Current capabilities
The v4 implementation contains 17 major capability areas:
| # | Capability | Implementation focus |
|---|---|---|
| 1 | Passive learning | Agents update behavior from simulated experience |
| 2 | Meta-learning | Evolvable learning-rule parameters |
| 3 | Darwinian evolution | Mutation, crossover, topology change, and speciation |
| 4 | Validated self-modification | Candidate changes are evaluated before retention |
| 5 | Self-directed control | Agents choose actions from internal state and goals |
| 6 | Self-evaluation | Curriculum, diversity, and stagnation signals |
| 7 | Recovery mechanisms | Rollback and anomaly-handling paths |
| 8 | Intrinsic exploration | Curiosity, novelty search, and self-play-inspired signals |
| 9 | Open-ended behavior search | Evolution can discover unprogrammed behavior combinations |
| 10 | Decision intelligence | Causal memory, prediction, and temporal evaluation |
| 11 | Emergent communication | Evolvable signaling between nearby agents |
| 12 | Counterfactual reasoning | Alternative-action replay and regret-based adjustment |
| 13 | Persistent knowledge transfer | Shared knowledge survives individual agents |
| 14 | Hierarchical goal formation | Multi-level goals and sub-goal decomposition |
| 15 | Theory-of-mind approximation | Internal prediction models for other agents |
| 16 | Abstract concept formation | Prototype-based compression of repeated situations |
| 17 | Self-narrative | Compressed autobiographical state influencing later decisions |
What changed in v4.0
v4 adds four capability families on top of the v3 communication, counterfactual, and shared-knowledge systems:
- Hierarchical goal formation — agents can maintain goals and decompose them into smaller objectives.
- Social prediction — agents model aspects of other agents' behavior to influence cooperation and competition.
- Abstract concept formation — repeated experiences can be compressed into higher-level prototypes.
- Self-narrative — agents maintain a compact history that can affect later decisions and inheritance.
Requirements
- Python 3.11 or 3.12 recommended
- NumPy
- Matplotlib
Install dependencies from the repository manifest:
python -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install -r requirements.txt
On Windows PowerShell, activate the environment with:
.venv\Scripts\Activate.ps1
Run
python genesis.py
The default configuration currently runs 150 generations with 100 simulation steps per generation, so a full run is intentionally more substantial than a smoke test.
Verify core behavior
Run the deterministic fast test suite without starting the full simulation:
python -m unittest discover -s tests -p "test_*.py" -v
The current core tests verify innovation-ID stability, minimal genome topology, finite bounded network activation, structural independence after genome copying, and learning-rule weight bounds.
Output
Generated plots are written under genesis_output/. The v4 visualization paths currently include:
genesis_output/genesis_v4_dashboard.pnggenesis_output/genesis_v4_universe.png
Generated output and Python cache files are ignored by Git so experiments do not continuously add local artifacts to source control.
Continuous integration
The GENESIS quality workflow performs fast checks on pull requests and pushes that touch GENESIS:
- compiles
genesis.pyand the core tests on Python 3.11 and 3.12; - installs the declared runtime dependencies;
- verifies that NumPy and Matplotlib import successfully;
- executes the deterministic GENESIS core behavioral test suite.
The workflow intentionally avoids running the full 150-generation simulation on every commit. Long experiment runs should be executed separately and their parameters/results recorded explicitly when used as evidence.
Architecture
See GENESIS_Architecture.md for the extended architecture notes. Where that document and the executable source disagree, treat genesis.py as the current implementation and open an issue or pull request to synchronize the documentation.
Reproducible research guidance
For comparable experiment results, record at minimum:
- the Git commit SHA;
- Python version and dependency versions;
- random seed, when fixed;
- configuration changes relative to
Config; - number of generations and steps per generation;
- the metric definition used for every reported result.
This separates observed experiment results from capability descriptions and makes future improvements easier to validate.
GENESIS v4.0 — an experimental platform for studying evolutionary, adaptive, social, and self-evaluating agent mechanisms.