The Distributed Artificial Intelligence is a subfield of Artificial Intelligence(AI) devoted to researching and developing the distributed solutions. Distributed Artificial Intelligence emerged in the 90s as a subfield of AI that deals with the interaction of intelligent and distributed agents. But it is getting much attention in recent times because of the decentralized and distributed nature of this field.
To make Artificial Intelligence more reachable and scalable, AI needs to be distributed and decentralized. Distributed AI systems are recently collaborating with Machine learning and Deep Learning Technologies, giving them a new dimension to explore. Today we will be talking about the latest trends in Distributed Artificial intelligence.
The rapid rise of the applications powered by Artificial Intelligence raises the data center's technical requirements, which generates high costs. The cloud is not a valid alternative in many cases. Artificial Intelligence's Future - Edge AI
The inspiration for swarm intelligence comes from nature. We have seen in our environment how bees distribute their work among themselves. Similarly, it amazes one how the flock of the birds communicates while flying in groups of more than a hundred. A similar pattern is also visible in the ants’ colonies. These all show the wisdom of crowd intelligence.
Swarm or collective intelligence consists of multiple agents(autonomous entities performing the task) that are decentralized and capable of self-organizing. The term was first coined by Bloom(1995) as a result of research on complex adaptive systems.
How does Swarm Intelligence Work?
Naturally, swarm intelligence happens through a process called stigmergy. It is a kind of natural mechanism for the coordination among the agents. When a task is performed in the environment, a trace is left intentionally by the agent. The leftover trance then triggers another event. In this manner, the whole series of tasks are performed until the defined goal is achieved.
This section will discuss some of the trending optimizing algorithms in swarm intelligence; these optimization strategies have made it the area of interest for researchers. Let’s see what amazes them:
Particle Swarm Optimization
Inspiration of the technology: It is inspired by the social foraging behavior of some organisms. Like the flocking behavior of birds and the schooling behavior of fishes. The way these organisms pass information while moving in the right directions in these swarms gives researchers opportunities to develop such systems artificially.
Methodology: The main objective of this algorithm is to make use of all agents to locate the optima in a multidimensional space. This optimum is initially assigned with any random position and velocity in the space, but as time elapses with exploration and exploitation, the optima are found. For greater understanding, see the image below.
Figure1. Flow of PSO
Applications: The applications of Swarm Optimization are listed below.
Swarm-based dimensionality reduction is used for dimensionality reduction in machine learning. It uses the vectorized implementation of the PSO.
As it is an optimization technique, PSO has also been used to hyperparameter tuning the deep learning algorithms.
Inspiration of Technology: It is inspired by the communication of the ants, which is done by using a harmonic chemical known as a pheromone. The agent’s probability of choosing the path is a function of the chemical intensity and the distance between the location. This phenomenon of using pheromone for communication is one of the types of stigmergy. Researchers have developed a similar artificial structure for this phenomenon.
Methodology: This strategy aims to use historical information and construct the solution for the individual agent using a probabilistic step-wise approach. The probability of selecting any component for constructing a solution depends on that component’s heuristic contribution to the overall cost function. Once the cost function is calculated, the history related to that path is also updated.
Applications: This Optimization technique is heavily used in a field called “Swarm Robotics,” which is the study about how the large number of simple robotic agents can be developed such that the desired goal can be achieved by the collective behavior of these agents using Ant colony optimization techniques.
Artificial Bees Colony Algorithms
Inspiration: It is inspired by the natural communication and distribution within the beehive. The scout bees are sent from the hive to locate the nectar. They return to the hive and give information about the hive's location, fitness, and distance of food using a waggle dance.
Artificial Bee Colony Algorithm can be used as the optimization function as an alternative to traditional gradient descent algorithms.
The Modified Version of the Artificial Bee Colony (ABC) can solve the clustering problem as an alternative to traditional clustering approaches.
Leading Players in Swarm Intelligence
Some notable organizations working in this field are Unanimous AI, Augur, Estimize, Almanis, Ace, etc. These are some of the leading players who believe in the idea of the crowd’s wisdom, i.e., swarm intelligence.
To understand Multi-agent, let’s first understand what an agent is in the case of MAS. An agent is a computer system capable of taking independent(autonomous) action to perform the specific task on behalf of its user.
A multi-agent system consists of several agents interacting with one another. Generally, agents will interact with other agents to fulfill their user's goals by cooperating, coordinating, and negotiating. The difference between the Multi-agent system and the swarm intelligence is that we have homogeneous agents in later ones, but we have heterogeneous agents in multi-agent systems.
What’s Trending in Multi-agent Systems?
The research in Multi-agent systems has its roots from the starting of distributed systems; with time, it has given rise to many sub-fields like game theory and negotiation theory. But with the increase in Machine learning and deep learning technologies has given MAS the opportunities to explore new dimensions. We will be talking about these new trends.
Multi-agent Reinforcement Learning
Multi-agent Reinforcement Learning (MARL) aims to develop multiple reinforcement learning agents (these agents will use deep learning while performing the task). These agents learn by dynamically interacting with the environment in which they are present. One can understand them as using enforcement learning(Machine learning technique) in a multi-agent setup.
MARL makes it possible for multiple agents to learn, collaborate, and interact collectively.
Applications of Multi-agent Reinforcement Learning
MARL is a new field of research let’s see some promising and fascinating potential use cases of MARL
It can be used in online resource allocation in a network of computing.
Research is going to develop applications of MARL in cellular Network optimization.
Last but not least, in the future, it can be used smart traffic control systems can be developed using MARL.
ABM can be thought of as part of the multi-agent systems as both have distributed autonomous agents designed to interact within an environment. Still, one crucial difference between them is that ABM is generally designed to simulate the system’s behavior. At the same time, MAS is used to solve a specific engineering problem.
Why to use Agent-Based Modelling?
Simulations model created Using ABM we can be used for the following purposes:
For a Depth Understanding of the system
For testing the coherence and comprehensiveness of the system
For Validation of theory against real data
For Making predictions
ABM in Working- Modeling Big Data of Organizations
Today every Organization have large volumes of data, and organizations are eager to get analyzed this data so that some fruitful insights can be taken out of it to make smart and intelligent decisions.
The data sources for this valuable information are CRM, ERP, HR, and other databases. ABM can be used to put that information to work and using simulation models; the insights can be delivered from the data, which can be used for forecasting, decision making, etc.
Traditionally, machine learning approaches are based on centralizing the training data in one place(a machine or a data center). But the idea of federated learning enables to allow the collaborative learning from distributed devices across the users.
How does Federated Learning work?
Federated Learning works like this. Each distributed device(say a smartphone) gets the current model. The model improves by learning the data on the device; the model generalizes itself and updates its weights and biases. The weights and biases are sent to the cloud in an encrypted manner after this cloud is updated with the weights coming from all the distributed devices, which updates the shared model. See the image below to understand the flow.
Figure2. Working of Federated learning
Smartphones update the weights locally according to personal data(A). Many users update it (B) Change is done on the cloud the shared model(C) the same is repeated again
Why Federated Learning?
Federated Learning enables the device’s data to remain on the device; hence, the user’s privacy is protected.
Federated Learning gives a favorable environment for the collaborative learning of the shared models rather than a generalized pretrained model; the model learns with time and with the collaborative effort of all distributed devices as agents.
It reduces the latency with significant factors as now the model remains on the device, and the dependency on the central cluster is reduced.
Hardware infrastructure dependency required to get the data in a single place is also reduced to a minimum level as the weights are sents that are the results of the training.
We have seen the latest trends in Distributed AI, though these trending technologies are in their infancy stage. The pace with which work is going in these domains, these all technology will make distributed Artificial intelligence feasible and state of the art solutions in the very near future.