The importance of Machine-to-Machine Economy (M2M) & Multi-Agent Systems

The Importance of Machine-to-Machine Economy (M2M) & Multi-Agent Systems

towards-data-science

This post was originally published by Alexandre Gonfalonieri at Towards Data Science

During my latest mission, I was in charge of developing a strategy related to decentralized artificial intelligence in the context of what we call “Machine to machine economy” (M2M). In this article, I’ll explain why Multi-agent systems are key to creating new business models in the upcoming M2M economy and why this new era could be a major threat to your organization.

Concretely, we expect machines to become “smarter” and soon be capable of making decisions and conducting transactions between themselves without any human interference. As a consequence, we will soon see new business models and customer relationships thanks to decentralized artificial intelligence.

Let us start by defining some key terms:

Agents: Sophisticated computer programs that act autonomously on behalf of their users, across open and distributed environments, to solve a growing number of complex problems. Increasingly, however, applications require multiple agents that can work together. (1)

Multi-agent system (MAS): A loosely coupled network of software agents that interact to solve problems that are beyond the individual capacities or knowledge of each problem solver. (2)

In simpler terms and using an example from Yoav Shoham and Kevin Leyton-Brown, you should imagine a personal software agent representing you on several e-commerce websites. For instance, let us assume that the task of this agent is to identify specific products available for sale in various online websites over time, and to purchase some of them on your behalf. In order to be successful, your agent will need to remember your preferences for products, your budget, and in general your knowledge about the environment in which it will operate.

Moreover, the agent will need to leverage your knowledge of other similar agents with which it will interact (in an auction, or agents representing other businesses).

A collection of such agents forms a multi-agent system. To elaborate, a Multi-Agent System is “a loose ecosystem of various communicating Artificial Intelligences” (3). It is essentially the next iteration of agent-based systems. Some algorithms have proved to be quite interesting in the development of MAS (reinforcement learning, deep learning, deep convolutional networks, …).

Currently, MAS (in a decentralized artificial intelligence context) is still being researched. As such, the industrial application/scalability of a multi-agent system is still a few years away.

A multi-agent system falls into one of two categories:

As we have seen, agents can interact with each other. The communication or coordination between such agents can take many forms. However, it is also necessary to understand that they are also autonomous agents. As such, in several cases, agents have opposite objectives and, therefore, are not able to carry out any cooperative process that includes them.

It is key to know that agents are equipped with a social capacity. This capacity can be defined by the “ability to exchange high-level messages (and not only data-bytes without an associated meaning) and carry out processes of social interaction with other agents (and/or humans) similar to those used by humans in their daily lives, establishing collective behaviours.” (4)

Our current challenge is to build agents that can negotiate and cooperate with other agents. For example, in order to convince an agent to cooperate, it might be necessary to make a payment or offer a particular good or service.

Moreover, we are also paying attention to Federated Learning.

Federated learning: A machine learning technique that trains an algorithm across multiple decentralized edge devices or servers holding local data samples, without exchanging them. (5)

Indeed, we believe that AI and real-time data processing must occur on edge networks and edge devices (IoT devices, etc). Is it key for some products to collect data from sensors to make decisions in real-time with no dependency on the cloud or internet. They should also be able to learn or train themselves using algorithms onboard and share their learnings with other products.

With the rise of AI, smart sensors, decentralized P2P transactional protocols, and Blockchain, we are moving into a machine economy. Machines will eventually become economically independent. An object (ex: car) will be able to spend money and earn money. The machines will make decisions for us, which explains the AI aspect, but we need to trust them — which is where blockchain comes in.

Our future customers will be machines with wallets.

We envision a future in which machines can discover, automatically connect with other machines (using public or private networks, for instance, a mesh network), make their own choices thanks to AI (Reinforcement Learning), and independently purchase what they need. We assume that machines will sooner or later have integrated wallets too.

An IoT device will no longer be considered as an isolated product that needs to learn everything from scratch on its own; these devices will leverage the mass learning learned by other similar IoT devices worldwide as well. This means that intelligent systems of IoT trained by machine learning are not just becoming smarter; they are getting smarter faster over time in exponential trends.

Moreover, this M2M economy will also be the one in which most customers will be charged by how much they use or consume a product. This is a major shift in the way we purchase products.

In the automotive industry, we could imagine vehicles that can seamlessly connect and communicate (using Multi-Agent Deep Reinforcement Learning) with other vehicles, roads, traffic lights, parking meters, gas pumps and even private companies such as Uber. More broadly, we can envision a society in which cars, drones, or buildings negotiate directly with each other to achieve their objectives without the necessity for human involvement.

To create new machine centered business models in the machine-to-machine economy, we must first improve our AIoT strategy by better using data network effects.

Machines are the customer of the future …

From a technical perspective, building a truly autonomous and scalable machine with the ability to make decisions beyond the context of a very specific purpose is still something extremely difficult.

The transformation of our products follows, more or less, the same roadmap established by the Commonwealth Bank of Australia. We three necessary steps in the evolution of the machine-to-machine economy

Transitioning into an economy in which most customers are machines will have a major impact.

  • How do you convince an AI agent to purchase from your company?
  • How do you manage “customer” loyalty in the M2M economy?
  • How can you ensure to remain relevant in this machine powered economy?

This situation will create new dynamics in your industry. Today’s customers who will rely on AI agents might shift to more “AI-friendly” businesses or some companies might create specific exclusive environments. At the same time, you must take into consideration that your products might independently generate additional revenue.

For instance, an autonomous car owner, in the M2M economy, can benefit from providing rides to other people or by selling or renting data sets to AI companies. One of our projects is to think about an incentive for this case. Indeed, as an incentive, the car could receive bitcoin or some tokens.

We are already concerned with the concept of trust between agents. Indeed, trust can be seen as the quality and quantity of interactions among agents: the more interactions occur between two parts, the more one trusts the other. Would it be easy to change this aspect in a competitive market driven by decentralized AI?

This issue of trust between agents can be solved by the Dynamic Interaction Based Reputation Model (DIB-RM) that was introduced to capture this dynamic property of trust. This model computes a reputation value for each agent on the system combining different dynamic factors. The reputation value is updated at each interaction.

Beyond obvious technical issues, companies must also transition from siloed solutions to a shared and trustworthy method of communication. Indeed, MAS can only be feasible for all industries if special protocols are developed for it. It will be hard for MAS to work under current data-driven protocols.

Spread the word

This post was originally published by Alexandre Gonfalonieri at Towards Data Science

Related posts