Evolution of Fear and Social Rewards in Prey-Predator Relationship

Technical Summary: Evolution of Fear and Social Rewards in Prey-Predator Relationship

1. Problem Statement

While it is widely accepted that fear is a fundamental evolutionary mechanism for avoiding danger, the specific evolutionary pressures that shape the complexity of fear and its interaction with other reward systems (such as social rewards) remain poorly understood. Specifically, researchers seek to understand:

• How varying levels of predatory pressure influence the evolution of fear.
• How predation interacts with the evolution of sociality (grouping behavior).
• Whether fear evolves in response to static environmental dangers (like pitfalls) in the same way it evolves in response to active, hunting predators.

2. Methodology

The authors developed a distributed evolutionary simulation framework using Reinforcement Learning (RL) agents in a 2D physics environment.

• Agent Architecture: Both prey and predators are controlled by three-layer Multi-Layer Perceptrons (MLPs). They use Proximal Policy Optimization (PPO) to learn motor policies based on sensory inputs (proximity sensors, tactile sensors, velocity, etc.).
• Evolutionary Mechanism: The simulation employs a stochastic birth-and-death model driven by energy metabolism. Agents inherit reward functions (genotypes) from their parents. These reward functions are composed of weights ($w$) assigned to different stimuli: food intake, motor activity, observing prey, and observing predators.
• Species Dynamics:
  • Prey: Consume food to gain energy; must avoid predators.
  • Predators: Hunt prey to gain energy; cannot eat food.
• Experimental Variables:
  • Predatory Pressure: Varied by changing the predator's "mouth size" (hunting capability).
  • Resource Availability: Varied by changing the food regeneration rate ($\Delta n$).
  • Control Environment: A "pitfall" environment where agents face static, non-lethal obstacles instead of active predators.

3. Key Contributions

• Co-evolutionary Modeling of Rewards: Unlike previous studies that focus on physical evolution or fixed reward structures, this work allows the internal motivation (the reward function itself) to evolve.
• Identification of Evolutionary Branching: The study demonstrates that a single population can split into distinct phylogenetic groups with different survival strategies (e.g., social grouping vs. individualistic escape).
• Distinction Between Active and Static Threats: The research provides evidence that the "active" nature of predators (movement and population fluctuations) is a primary driver for the evolution of fear, which is not replicated by static environmental hazards.

4. Results

• Emergence of Fear and Sociality: In default settings, prey evolved negative rewards for observing predators (fear). However, a significant finding was the emergence of social rewards (positive rewards for observing other prey). In some lineages, prey evolved to prioritize grouping as a defense mechanism, even if it meant a reduced individual fear response.
• Predatory Pressure vs. Fear: Increasing predator hunting capability (larger mouth size) directly intensified the evolution of fear (more negative $w_{pred}$) and social rewards (higher $w_{prey}$) in prey.
• Resource Scarcity vs. Sociality: When food was scarce, prey evolved to be more solitary (negative $w_{prey}$), suggesting that the need for individual foraging outweighs the benefits of social grouping in resource-poor environments.
• The Pitfall Control: In the pitfall environment, agents did not evolve fear. Instead, they evolved a positive reward for approaching pitfalls ($w_{pit} > 0$) because food tended to cluster near these dangerous areas. This suggests that if a threat is static and non-lethal, agents will prioritize "risky foraging" over avoidance.

5. Significance

This paper contributes to both evolutionary biology and artificial intelligence. By showing that complex reward structures—including fear and sociality—emerge from the interplay of predation and resource competition, the study provides a biologically plausible model for the evolution of emotions. For the AI community, it demonstrates that multi-agent evolutionary simulations can move beyond simple "optimal decision-making" to explain the emergence of complex, diverse, and sometimes counter-intuitive behavioral strategies.