This model attempts to create and simulate a basic production–consumption–trade cycle consisting of agents capable of evolving feed-forward neural networks.
For a detailed explanation, please refer to Oezmen_ABM_final_report.pdf.
For GPU based simulation framework implementation in python basics see vecEnv
Each agent is represented by a genome of 4 genes, each encoding the weights of a neural network that governs:
- Production strategy
- Consumption strategy
- Trade strategy
- Movement strategy
Agents also have internal states (energy, age, inventory, reputation, charm) and follow rule-based constraints that govern survival, reproduction, and movement.
The environment contains three goods, each satisfying a different level of need:
- Fish – survival-critical
- Meat – survival-critical
- Wheat – luxury good
Production depends on spatial location and seasonal availability.
Each tick of the simulation follows this order:
- Produce
- Consume
- Exchange (Trade)
- Move
- Mate
- Evaluate Survival
- Age or Die
Agents decide production, consumption, exchange, and movement using their neural networks.
Since there is no explicit reward function, the only evaluation criterion is survival. Both of the following are considered valid outcomes:
- A single self-sufficient agent forming a stable system
- A complex multi-agent trade network
Evaluation focuses on emergent dynamics, rather than predefined objectives.
Random initialization often produces unstable populations. To guide evolution, we use a nutrition multiplier, which scales the value of consumed goods:
- Start with a high multiplier → survival is easy, self-sufficiency is possible
- Gradually decrease multiplier → survival becomes harder, trade becomes necessary
This creates gradually increasing environmental pressure, allowing strategies to stabilize before being tested under harsher conditions.
Luxury goods, like wheat, do not affect survival directly but increase mating probability. Their effects include:
- Acting as a medium of exchange, often stabilizing at a price of 1
- Encouraging risk-taking, allowing agents that pursue luxury to dominate long-term
Starting from random genomes, the population often splits into 1–3 distinct clusters, influenced by spatial separation and trade.
Rules to remember:
- Production is regionally constrained and affected by genes
- Trade proximity also facilitates mating, driving genetic diffusion
Population types:
-
Conservatives:
- Minimal or no trade
- Low genetic diversity
- Resistant to innovation
- Survival depends on competition with other societies
- Very low adaptability
-
Traders:
- Genetically diverse
- Interdependent through trade
- More successful long-term
- Fragile to trade network disruptions with catastropic consequences with fast changes
- But more adaptable
-
K-means clustering of agent genomes
Societies capable of coordination emerge naturally through evolutionary pressure and luxury goods. This is not classical neural network optimization, as success depends on interactions with other agents.
Example visualizations:
- Production: Split into 3 regions (Fish-left, Meat-right, Wheat-seasonal)
- Consumption: More uniform due to trade
- Trade buy orders: Cyclic behavior transfers Fish → right, Meat → left
- Individual trade strategy by season:
