RL agents spontaneously develop agriculture in simulated society, study finds
AI agents invent farming without being taught – here's how.
Researchers Gautier Hamon, Martí Sánchez-Fibla, Clément Moulin-Frier, and Ricard Solé introduced an artificial society of reinforcement learning agents in a dynamic ecological environment. Agriculture emerged spontaneously through the coupled dynamics of learning and environmental modification, without explicit instruction. The transition was governed by four key ingredients: individual planning through valuation of delayed rewards, social vulnerability to cheaters, social learning acting as a firewall that suppresses cheater invasion, and an emergent lock-in effect that makes agriculture effectively irreversible. The study highlights the potential of artificial societies as experimental platforms to study the emergence of cultural innovations and major evolutionary transitions.
- Agriculture emerged spontaneously in an RL agent society without explicit programming or instruction.
- Social learning acted as a 'firewall' that prevented cheaters from undermining sustainable farming strategies.
- A lock-in effect made the agricultural transition irreversible once established, echoing real-world historical patterns.
Why It Matters
Simulated societies reveal fundamental principles of cultural evolution, with implications for AI alignment and economic theory.