Our Core Vision and First Steps
In the realm of AI, intelligent agents are defined as systems that help achieve objectives set by users. These agents, or personal assistants, are usually adaptable, semi-autonomous software modules that can collaborate with users. Naturally, the evolution of hardware to software is succeeded by agentware, which characterizes the new architectural principles of information processing based on agents. The software technology of agents and agent-oriented programming are understood in general as a natural development of the ideas of object-oriented programming. Agents may be considered as active objects endowed with a certain degree of freedom that can manipulate other objects, create and destroy them, and interact with the environment and other agents. The fact that the same definition applies to smart contracts on the blockchain is what motivated the creation of Cerebrum. Cerebrum is attempting to transform smart contracts, which already behave as simple rule-based, autonomous agents, into data-driven intelligent agents that can be trained by leveraging the economic incentives that blockchain technology can provide. These economic incentives will harness the power of crowdsourcing in the context of machine learning and data sharing. The founders of Cerebrum believe this new category of blockchain-based, intelligent agents will be able to solve interesting problems in unique ways not possible without the advantages of blockchain and the economic incentives it can create.
Multi-Agent Systems
In classical AI, an agent, having a global vision of the problem, has all the necessary skills, knowledge and resources for individually solving a task. In contrast, multi-agent systems (MAS) assume that a single agent can have only a partial understanding of the task and can solve only some of its subtasks. MAS are used to solve complex problems that are difficult to solve by an individual agent and require interacting agents. The activity and organization of artificial systems and their collaborative approach to the concerted solution of tasks are fundamental characteristics of advanced information technology and network organizations, built on principles of MAS. Synergy within MAS is based on processes of interactions between collective agents, leading to the formation of artificial communities with fundamentally new features. The tasks in MAS are distributed between the agents, each of which is considered as a member of the organization. Distribution of tasks involves assigning roles to each of the agents. Depending on whether the distribution comes from the task set or the ability of each agent, one can distinguish between systems of distributed AI and systems of decentralized AI.
In distributed AI, the process that decomposes original tasks and combines task outputs to obtain solutions is centralized. MAS is rigidly projected downward on the basis of partitioning the general task into separate, relatively independent subtasks and preliminary determination of the agents’ roles. In decentralized AI, the distribution of tasks happens largely spontaneously, directly in process of the interaction and communication between the agents.
Distributed AI requires the development of organizations capable of solving tasks in unison. These organizations consist of a set of individual agents, each with its well-defined subtask. Distributed AI systems exhibit three main characteristics: distribution of tasks, distribution of power, and communication between agents. The typical scheme of a MAS based distributed solution for a task includes these steps:
1. Task decomposition — decompose the original problem into separate subtasks
2. Subtask distribution — distribute the subtasks to various agents that solve or divide them
3. Subtask composition — combine the results from each of the subtasks for the overall result
The ideology of distributed task solving assumes the separation of knowledge and resources between the agents and distribution of management and power. Ideally, a governing body provides a common model with global criteria for achieving goals.
In fully decentralized systems, management takes place only because of local interactions between agents, not a distributed solution of some general task, but a coordination of autonomous agents in a dynamic multi-agent world. Local tasks of individual agents, solved on the basis of local models and criteria, are described along with the distributed knowledge and resources.
Distributed AI can solve a range of tasks, but it is strictly centralized and limited by design pattern. Decentralized AI removes points of failure and creates a trustless, autonomous environment, but it is limited in goal complexity and adaptability. Combining distributed AI with decentralized AI into a new AI system based on existing models of multi-agent systems compensates their shortcomings while enhancing their benefits.
The Cerebrum AGI
Cerebrum will realize both distributed and decentralized AI by using the MAGIC protocol for distribution and integration of tasks, Ethereum smart contracts for implementing autonomous agents, and IPFS for data decentralization. Multi-agent systems communication provides a way for multiple interacting intelligent agents to communicate with each other and with their environment. Cerebrum will employ the Multi-Agent Generated Information Communication (MAGIC) protocol to allow communication between agents. The MAGIC protocol enables hosts to coordinate the transfer of outputs of various agents to any number of other agents. As a result, higher-level competitions can be configured to take the best outputs of competitions and leverage them as host datasets for scientists to further model. All interactions between agents are handled autonomously via CNS and Cortex. Thus, Cerebrum is protected from any risks associated with a central authority. Cerebrum combines decentralized, distributed general-purpose tasks with the MAGIC protocol and smart contract based rewards, powered by the Ethereum blockchain.
Consensus over the definition of true artificial general intelligence (AGI) is not well-established, however, several key concepts resonate through AGI research: an AGI should be autonomous, goal-directed, highly adaptive by utilizing self-learning mechanisms, and good at generalizing across various goals and contexts by stressing on the general-purpose nature of intelligence. In the Strong Mind state, Cerebrum will launch a blockchain-based AGI that creates agent combinations and uses feedback mechanisms to optimize those combinations and usage of various agents for specific tasks. Once the AGI learns to optimize multi-agent configurations it can be updated to a general-purpose distributed intelligence, which acts as an autonomous controller that can divide tasks by constructing new agents leading to a multi-purpose AI that solves arbitrary goals. The Cerebrum AGI will rely on the premise of an artificial organism, an intelligent agent that exists in a population of its kind and tends to learn, adapt, and defeat competitors in the environment to survive. This agent-model is based on artificial evolution whereby mutations of agents persist through survival-of-the-fittest and based on principles of artificial life revolving around self-reproduction and self-preservation.
Join our community on Twitter, Slack, Reddit, Facebook.
Learn more about Cerebrum at https://cerebrum.world.
Acknowledgements:
