A study of evolutionary multiagent models based on symbiosis

Area of Science:

• Artificial Intelligence
• Computational Intelligence
• Evolutionary Computation

Background:

• Conventional Multiagent Systems (MAS) often focus on individual agent benefits, potentially leading to suboptimal outcomes like Nash Equilibria.
• Symbiosis, a concept from ecology, offers a novel framework for inter-agent interactions.
• The need for MAS methodologies that promote cooperation and achieve superior performance is recognized.

Purpose of the Study:

• To investigate the evolutionary model of Multiagent Systems with Symbiotic Learning and Evolution (Masbiole).
• To analyze agent behaviors resulting from symbiotic evolution within the Masbiole framework.
• To demonstrate the advantages of Masbiole over conventional MAS in specific simulated environments.

Main Methods:

• Development and application of the "Match Type Tile-world" (MTT) simulation environment, adapted for symbiotic agent interactions.
• Utilizing "Genetic Network Programming" (GNP), a novel evolutionary computation method with a directed graph gene structure, for analyzing agent decision-making.
• Simulating agent evolution based on symbiotic learning and evolution principles, where agents consider mutual benefits and losses.

Main Results:

• Masbiole demonstrates the ability to escape Nash Equilibria, a common limitation in traditional MAS.
• Agents evolved through Masbiole exhibit diverse and effective behaviors.
• Simulation results in the MTT environment show that Masbiole achieves better performance compared to conventional MAS.

Conclusions:

• Symbiotic evolution in Masbiole facilitates the development of agents capable of complex, cooperative behaviors.
• Masbiole offers a promising alternative to conventional MAS, yielding enhanced performance through consideration of symbiotic relationships.
• The combination of MTT and GNP provides an effective platform for analyzing and understanding symbiotic agent evolution.