TLDR: This research explores how complex behaviors like resource gathering and predation emerge in large-scale, unsupervised ecological simulations. Agents evolve through natural selection without explicit rewards, interacting with their environment. The study found that larger environments and populations, along with advanced sensors like a compass and vision, are crucial for the consistent and stable emergence of sophisticated behaviors, suggesting that ecological scale can drive the development of intelligence.
In the quest to understand how complex behaviors arise in nature, scientists often face immense challenges. Natural ecosystems are vast, intricate, and difficult to control or measure without risking harm to wild populations. A new research paper titled, “THE EMERGENCE OF COMPLEX BEHAVIOR IN LARGE-SCALE ECOLOGICAL ENVIRONMENTS,” by Joseph Bejjani, Chase Van Amburg, Chengrui Wang, Chloe Huangyuan Su, Sarah M. Pratt, Yasin Mazloumi, Naeem Khoshnevis, Sham M. Kakade, Kiant´e Brantley, and Aaron Walsman, explores a novel approach: simulating evolution in massive digital worlds.
This study delves into how the physical size of an environment and the sheer number of inhabitants influence the development of sophisticated behaviors. Unlike traditional machine learning where agents are given specific goals or rewards, here, agents are unsupervised. They simply exist, evolving over time through processes mirroring natural selection: reproduction, mutation, and survival of the fittest. As these digital organisms act, they continuously reshape their environment and the populations around them, creating a dynamic and ever-changing ecology.
The primary goal of this research isn’t to create a single, highly optimized artificial intelligence. Instead, it’s to observe and understand how diverse behaviors naturally emerge and evolve across large populations under the pressures of competition and survival. To achieve this, the researchers conducted experiments in truly large-scale digital worlds, some hosting over 60,000 individual agents, each powered by its own evolved neural network.
The experiments revealed several fascinating emergent behaviors. These included agents developing strategies for long-range resource extraction, using their vision to forage more effectively, and even engaging in predation. A key finding was that these complex behaviors often only appeared in sufficiently large environments and populations. Furthermore, larger scales tended to make these behaviors more stable and consistent over time.
For instance, agents equipped with a simple compass sensor demonstrated the ability to undertake long-distance journeys inland to gather resources, a behavior the researchers termed “mining.” They would navigate from water to land, collect biomass left by deceased ancestors, and successfully return to the water. This behavior was far more consistent and effective in larger environments. Without a compass, agents often perished on land, leading to a resource imbalance and eventual population decline.
Similarly, agents with vision sensors (a 7×7 top-down view) evolved superior foraging and attacking skills compared to those without. Visual agents spent less time moving and more time eating, leading to larger and more stable populations with better resource utilization. In terms of predation, visual agents attacked less frequently but with significantly higher precision, suggesting the emergence of specialized hunting strategies. These advantages of vision also became more pronounced and stable in larger environments.
The simulations were built using a new JAX-based environment, allowing for rapid evaluation of large grid worlds. These environments featured varied terrains like oceans, beaches, islands, lakes, isthmuses, and channels, ranging in size from 64×64 to 1024×1024 grid cells. Agents needed to collect water, energy, and biomass to survive and reproduce. Their actions included resting, moving, attacking, and eating, with policies implemented as small, memoryless neural networks that mutated upon reproduction.
Also Read:
- EvolveR: How AI Agents Learn and Grow from Their Own Actions
- Simulating Online Communities: When AI Agents Learn to Connect
This work suggests that open-ended evolutionary settings offer a powerful new paradigm for studying the emergence of intelligence. By combining ecological and evolutionary pressures with rich, large-scale environments, we may find a natural pathway for developing increasingly intelligent artificial agents. The experimental code for this research is available at https://github.com/jbejjani2022/ecological-emergent-behavior.
