Why Colony of AI?

Author(s): Shan Suthaharan PhD

Originally published on Towards AI.

Real-world applications are generally composed of multiple, related and unrelated, problems. These problems relate to parallel, hierarchical, and sequential structures, forming a single complex, heterogeneous, and scalable problem. It requires multiple tasks that align with these parallel, hierarchical, and sequential properties of the problem. Hence, real-world applications, consisting of multitask problems, require multitask solutions. We can name a few such multitask applications: disaster response and emergency management, healthcare and pandemic management, and climate monitoring and environmental protection. Since the problems in these multitask real-world applications are connected, the tasks must be carried out with coordinated efforts and knowledge sharing. In addition, the rapid changes in the problems over time must also be taken into consideration when implementing multitask solutions.

Multi-agent AI solutions

Due to the complexity of the inherent problems in real-world applications, artificial intelligence (AI) system would be a suitable mechanism to develop multitask solutions [1]. To develop such AI solutions for multitask applications, multi-agent AI systems that can work in parallel with coordination are required. However, the multitask real-world applications have many demanding requirements that may not be able to be met by the traditional multi-agent AI systems [2]. For multi-agent AI systems to be suitable for a wide range of real-world multitask applications, they need to satisfy several requirements. The revolutionary colony of AI framework can satisfy such requirements [3, 4]. These requirements are discussed in detail in this article, by highlighting key features of the concept of colony of AI, to answer the question of “why colony of AI?” in a systematic manner.

Requirement 1: Intrinsic Coordination

In multitask applications, a task may depend on the results of some or all the other tasks. Hence, AI agents working on different tasks or the same tasks must perform the tasks collaboratively and share the learned knowledge timely. This is the task-oriented intrinsic coordination that is required by multitask problems. The integration of nature-inspired concepts in the colony of AI can assist AI agents to coordinate efficiently and share their learned knowledge (weight and biases) and intelligence by mimicking the behavior of biological systems, like ant and bee colonies.

In traditional multi-agent AI systems, the coordination efforts are generally limited and require user interaction. In other words, the coordination strategies must be predefined and integrated into AI agents. But in the colony of AI systems, the coordination efforts are enabled by the integrated AI-driven strategies that direct AI agents to independently acquire relevant knowledge from local features, while developing evolutionary intelligence and contributing to the collective decisions of the AI colony.

Requirement 2: Diversified Learning

To facilitate efficient coordination scheme for multitask real-world applications with multi-agent AI system, some tasks must be completed earlier than the other tasks, based on the dependency structure of the tasks. Additionally, the heterogeneous structure of the problems of these applications requires diverse learning. In essence, multitask applications require reliable AI agents that learn the tasks fast, detailed and organized modes. But the traditional multi-agent AI systems are generally rule-based systems; thus, they rarely meet these requirements. As we can understand, real-world multitask applications require role-based systems where AI agents can learn in fast, detailed, and organized modes.

In other words, for example, AI agents in multi-agent AI systems must first develop skills to learn fast and provide on-demand approximate solutions for timely actions, while continuously learning to gain other forms of knowledge acquisition and provide endurance solutions that are stable and reliable. Hence, the colony of AI frameworks provide role-based fast, detailed, and organized learner AI agents that contribute to trustworthy and explainable decision-making through solution diversity and quality.

Multitask applications require solutions that can be trusted so that they can address the presence of complexity and uncertainty in the compilation of the problems. To trust the solutions from an AI agent, the AI agent needs to learn in detail and in depth. The colony of AI supplies detailed learner AI agents for this purpose. With the absence of such dedicated learning capabilities, along with their rigid structures, the current multi-agent AI systems cannot guarantee trustworthy and explainable solutions.

In addition to the fast and detailed learning strategies, multitask real-world applications require systematic learning strategies to overcome the problems associated with the heterogeneity of the problems and their scalability. The problems in real-world applications, as stated earlier, form parallel, hierarchical, and sequential complexities. Hence, the colony of AI provides organized learner AI agents that organize their learning strategies by following the same organized structure to address this requirement.

Fast learner AI agents can learn in the absence of some information; hence, they may not need to wait for the results from other AI agents; at the same time, the other fast learner AI agents also learn fast and arrive at some results and supply that information. This is an innovative coordinated logic structure that is unique to colony of AI. In essence, such interacting layers of coordination designs are also available among detailed and organized learner AI agents with respect to the need for detailed and organized decision-making with meaningful delays.

Requirement 3: Resilient System

Multitask applications require a multi-agent AI system that can recover fast if an AI agent working on a task fails. It requires the availability of a similar AI agent for immediate replacement of the failed AI agents for the system to recover faster. It would be even better if we had a system that could provide a feature that prevents (or minimizes) the system from failure.

The colony of AI provides multi-model and mixture-model concepts to establish multiple models of the same role (fast, detailed or organized) that can prevent the system from failing. Such resilience features are limited in the traditional multi-agent AI systems since their design structures are static and rigid because of the stricter rule-based multitask environment.

To generate multi-models and their families, the colony of AI introduced “triplets” structure to produce multiple AI agents of the same type that perform similar tasks. Similarly, to generate mixture-models and their families, the colony of AI introduces “hierarchical triplets” to produce multiple AI agents of distinct types that can perform dissimilar tasks.

Requirement 4: Evolutionary Intelligence

The problems in multitask real-world applications are evolutionary in the sense that any two or more problems interact and generate a new problem that can be undetected by multitask strategies. Therefore, it is important to observe parallel, hierarchical, and sequential structures of the inherent problems and identify evolved problems. To facilitate this, the colony of AI introduced the concept of marriage of AI agents for knowledge sharing by using genetic algorithms with crossover and mutation operations.

This mechanism is not available in the traditional multi-agent AI systems. Marriage of AI agents allows two AI agents to become parent-AI agents and produce child-AI agents by sharing their learned knowledge (weights and biases) selectively and transferring shared knowledge to child-AI agents.

Requirement 5: Explicit Localization

In many real-world applications, problems start at the local level and then spread to a global level. This is our natural understanding in life too. Hence, it is important to localize multitask solutions first and then develop a deeper understanding of the problems and their associated solutions. The colony of AI localizes the knowledge acquisition and the development of local solutions through the concept of families of AI agents.

With the families of AI agents, the colony of AI generates multiple AI agents to work together with similar interests and skills; thus, the families of AI agents localize the solution and contribute to solution diversity and quality decision-making. In colony of AI, the families of AI agents provide explicit localization with robust evolutionary intelligence that can help rapid recovery of the system. In contrast, the traditional multi-agent AI systems localize the learning and solutions to a single rule-based AI agent.

Requirement 6: Interaction Flexibility

In real-world applications, multiple tasks are performed to solve interacting problems that are inherent to these applications. Hence, these applications require flexibility of knowledge-propagation among the tasks. Strict rules can lead to destructive effect on the system when they are too rigid and inflexible, especially when the multitask problems are not fully understood. While multi-agent AI systems can provide some flexibility, the colony of AI uniquely defines a mechanism, called flexible parents, that is purely reserved to address bias management to impose flexibility.

In the process of marriage of AI agents (i.e., in the process of knowledge sharing), one of the parent-AI agents is considered flexible based on the understanding of bias, accordingly, modifies the weights (knowledge) and then shares the weights with the other parent-AI agent to produce child-AI agents. Also, note that colony of AI uses 50% of the weights for sharing. This strategy allows child-AI agents to understand the problem domain better.

Requirement 7: Heterogeneous Solutions

The parallel, hierarchical, and sequential structures of the problems in multitask real-world applications make the need for heterogeneous solutions as a requirement. They can utilize this diversity of solutions to establish useful connections between the heterogeneity of the problems and the diversity of the solutions.

The current research clearly shows that the traditional multi-agent AI systems do not directly incorporate mechanisms to generate diversified solutions. In contrast, the colony of AI, through its multi-model and mixture-model families of AI agents and evolutionary intelligence, generates solution diversity to address heterogeneity of the problems.

Requirement 8: Decision Globalization

Although real-world applications are composed of multiple problems, they prefer to generate single and final comprehensive solutions. In addition, they require robust solutions that are derived from the diversified local solutions. While the traditional multi-agent AI systems focus on generating such single solutions, they don’t guarantee the validity of the solutions, except trusting on the design structures of the systems. The colony of AI explicitly adopts a kernel density estimation technique (KDE) and delivers single and comprehensive solutions that are derived from the diversified solutions collectively generated through evolutionary intelligence concept.

Since the solutions are generated by many random processes throughout the development of a colony of AI framework, the system assumes randomness of solutions and applies kernel density function to estimate single KDE-based solutions; hence, it provides trustworthy solutions with explainability with the ability to tune the performance parameter “bandwidth” of the kernel density function. The absence of this quality is an observable disadvantage in the traditional multi-agent AI systems.

Requirement 9: Extended Memorization

The decision globalization and the evolutionary properties of the problems require an extended memorization of the models’ behaviors and solutions. In other words, multitask applications require that AI agents remember their local experience associating them with tasks to provide enduring solutions — multi-agent AI systems do not have such ability. The colony of AI introduced a mechanism to maintain family history in the colony. It uses a n-tuple structure that can preserve historical information of AI agents. This information includes the learner types of parent-AI agent, the data domain (or feature space) they are trained to acquire knowledge and intelligence, and the learner types of the child-AI agents. In addition, duration of training (epochs) and other performance parameters may also be preserved.

Major Highlights

Therefore, in summary, we can say that colony of AI delivers a unique multi-agent AI framework that exceeds the traditional multi-agent AI system by integrating the following special functionalities:

• Nature-inspired community of AI agents that mimic the behavior of biological colonies, like ant and bee colonies.
• Role-based learner AI agents — in addition to rule-based learner AI agents — that have fast, detailed, and organized learning abilities.
• Efficient coordination among AI agents with diverse multitask solutions and collective decisions that are validated by strong statistical tools.
• Marriage of AI agents that enable knowledge sharing among AI agents by using genetic algorithms with crossover and mutation operations.
• Families of parent and child AI agents that localize evolutionary intelligence and develop a fault-tolerant colony of AI agents.
• Triplets and hierarchical-triplets strategies that enable knowledge sharing between similar and dissimilar AI agents based on roles.
• A n-tuple framework to preserve family history, such as parent and child relationships, to develop long lasting solutions and decisions.

Therefore, overall, the colony of AI develops evolutionary intelligence among AI agents and delivers resilient multi-agent AI systems to address highly-complex multitask real-world problems and challenges.

Conclusion

We have seen significant differences between the traditional multi-agent AI systems and the revolutionized colony of AI frameworks. They first differ in their fundamental concepts—Multi-agent AI systems follow game-theoretical approaches, whereas the colony of AI frameworks follow nature-inspired approaches. Hence, the colony of AI brings all the advantages of coordination and collaboration characteristics of nature to develop solutions and address dynamical and evolutionary multitask problems. In contrast, learning is limited and generally restricted to individualized AI agents in traditional multi-agent AI systems. However, in colony of AI, learning is performed with shared efforts collectively among AI agents to generate evolutionary intelligence.

The multitask solutions are scalable like the solutions observable in biological systems, when the emergence of problems creates dynamic and evolutionary situations. The colony of AI also provides a framework that is flexible and modifiable to perform continuous research and advance multi-agent AI systems. In summary, the colony of AI delivers a computational framework that generates diversified solutions—opposed to the traditional multi-agent AI systems that provide fixed solutions—and high-quality decision-making with flexibility and explainability that lead to trustworthy AI systems. It is also important to note that in its present form the colony of AI focuses on computer vision problems with the potential to contribute to advance this research with natural language processing. In essence, its flexible and logical structures also make it a highly suitable multi-agent AI system that can solve multitask real-world problems, efficiently.

References

[1] Ferber, Jacques, and Gerhard Weiss. “Multi-agent systems: an introduction to distributed artificial intelligence.” Vol. 1. Reading: Addison-wesley, 1999.

[2] Luzolo, Pedro Hilario, Zeina Elrawashdeh, Igor Tchappi, Stéphane Galland, and Fatma Outay. “Combining multi-agent systems and Artificial Intelligence of Things: Technical challenges and gains.” Internet of Things (2024): 101364.

[3] Shan Suthaharan. “Colony of AI: Towards building families of AI-agents using theory of genetic algorithm and bias randomization.” In Emerging Topics in Artificial Intelligence 2024, vol. 13118, pp. 59–62. SPIE, 2024.

[4] Shan Suthaharan. “A nature-inspired colony of artificial intelligence Ssystem with fast, detailed, and organized learner agents for enhancing diversity and quality.” In Proceedings of the AAAI Symposium Series (Vol. 5, №1, pp. 431–438), 2025.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI