Social Insects and Network Load Balance

Social Insects - AI for Big Data

Introduction

The primary goal of Artificial Intelligence (AI) is to transform electronic brain depicted in Figure 1, into a human-like brain as in Figure 2, in order to gather data, process it and create models (hypothesis), predict or influence outcomes, by relying on huge data sets, which will improve human life.

Figure 1 - Human Brain Processing

Figure 2 - Electronic Brain Processing

The availability of Big Data has accelerated the growth and evolution of AI and machine learning applications. Data volumes are exploding; more data has been created in the past two years than in the entire history of the human race.
Here is a quick comparison of AI before and with with Big Data:

AI before Big Data	AI with Big Data
Availability of limited data sets (MBs)	Availability of ever-increasing data sets (TBs)
Limited sample sizes	Massive sample sizes resulting in increased model accuracy
Inability to analyze large data in milliseconds	Large data analysis in milliseconds
Batch oriented	Real-Time
Slow learning curve	Accelerated learning curve
Limited data sources	Heterogeneous and multiple data sources
Based on mostly structured data sets	Based on structured / unstructured and semi-structured data sets

The term Big Data represents growing volumes of data. Along with volume, the term also incorporates three more attributes, velocity, variety, and value:

1. Volume: The ever increasing and exponentially growing amount of data. Ex. a flight traveling from one point to another generates half a terabyte of data
2. Velocity: The amount of data generated with respect to time and the need to analyze that data in near real-time for some critical operations
3. Variety: The variety of data formats and structures. Not all data is structured in databases, especially social media data
4. Value: The data is only as valuable as its utilization in the generation of actionable insights

Note:

• In 1992: 100 GB / Day produced
• In 1997: 100 GB / Hour produced
• In 2002: 100 GB / Second produced
• In 2013: 28,000 GB / Second produced
• In 2018: 50,000 GB / Second produced

Swarm Intelligence

Ants Swarm - Photo by Poranimm Athithawatthee from Pexels

Ants work in a swarm mode, they move in a coordinated line one behind another. They can collect and carry foods much larger than their size all the way to their nests, and can form bridges to cover large gaps. 

Considering that their brains are nowhere close to the human brains in terms of neurons and connections, when they work in a group, they can achieve their goals intelligently; thus, they are called social insects. 

Those social creatures have prominent characteristics: 

• they live in colonies
• they have division of labor
• they have strong group interactions
• they are flexible

Collectively, they achieve intelligence, and this type of phenomena has prompted researchers to achieve what is called Swarm Intelligence (SI). 

The SI system is composed of a colony of agents (individual ants) which are also called boids. Each boid interact with its neighbor and environment (context) to achieve individual goals, and together, they achieve a larger goal without governance or centralized authority. 

In AI language, "Swarm intelligence is a collection of intelligent systems inspired by the collective intelligence of a group. This collective intelligence is achieved through the direct or indirect interactions of agents that are homogeneous in nature, yet co-operate with each other in their local environment without being aware of global context or pattern."

If you will build your SI-based system, there are three fundamental concepts that your system should at minimum comply with:

Self-Organization (SO)

SO is the property of SI systems that determines the underlying cooperation among SI agents to achieve a desired collective behavior. The agents are not aware of any global patterns or behavior, however, the global behavior is emergent out of individual functioning of agents.

Here's an example of self-organization: The ant colony as a whole is always striving to construct a nest that is safe from harsh environments and organize individual ant activities so as to locate the source of food that is nearest among all the available food sources. The ants apply a very unique and smart algorithm for locating the nearest and most abundant food source. Once the shelter (colony) is established, the most important aspect for the colony's survival is to find the nearest and most abundant source of food.

Stigmergy

The rules need to be reactive to the changes in the environmental state and the agent should be able to adapt to the changes autonomously and continue to perform its function.

Here's an example of stigmergy: an ant moving on a path to the food source and there is some water poured on the path. As soon as the ant encounters water on the way, it starts looking for an alternate path based on the pheromone(ant chemicals) signal. It may also traverse its way back to the colony and then start over again on another path autonomously (without any central control). At the same time, the ant leaves traces for other ants to know that on a particular path to the food source, there is trouble on the way. Other ants immediately adapt to the change in environment based on the previous ants' experience and modify their trajectories based on the simple rules. The ants interact with each other without any explicit communication, but only with the modifications in the environmental state.

Division of labor

The individual agent within the swarm is extremely limited in its capability to achieve the goal for the entire swarm. The natural system applies division of labor with individual agents performing a set of very specific responsibilities that contribute to the overall success of the swarm.

For example, all the bees in a hive are not doing the same thing. There is a clear division of labor within the bee hive based on the type of the bee. The Queen bee is responsible for laying eggs, the male drones are responsible for reproduction, and the worker bees build the hive and work to get food for the entire population. They also take care of the Queen bee and the drones by feeding them.

Applications in Big Data analytics - Network Load Balancing

In the current form of networks, the big data computing framework is an enormous collection of computation nodes that are distributed across the globe. 
Two types of deployment exists: on-premises and on the cloud.

Cloud is a virtualized infrastructure and it is geographically distributed in various regions where the nodes (computers) exist, and they are controlled by a centralized unit that keeps track of all nodes operations. 
However, the swarms do not have a central command and the agents work autonomously based on their rules, after which the agents adjust themselves to changes. 

SI concepts can be applied in securing the infrastructure as well as the nodes and making sure they are fully balanced. 

Without AI, we start by submitting a computation job to the master node, which in turn breaks it down into multiple chunks to be executed independently by the slave nodes, the jobs finish at different times and require different degrees of computation and storage. It may happen that the core compute load is not evenly distributed across the nodes. 

With AI, we deploy the Ant Colony Optimization (ACO) model in the distributed computing environment, and the process will be: 

Ant Colony Optimization - ACO

• Reproduction: the controller checks the platform periodically and generates artificial ants (agents) based on the load on the cluster nodes. If the nodes are overloaded or underloaded, new ants are generated for carrying the message across.
• Exploration: the agents are independently in charge of finding the nodes that are overloaded. They can trace the network operating parameters and leave a pheromone (incremental counter) for other swarm agents to get notified. 

How did the ants achieve load balance: 

• Agents calculate and quantify the (under or over) load at which it is connected
• Start in the direction of random node to check for load balancing suitability
• Backward artificial ant is generated when a candidate node is found. The agent updates the incremental counter for trail tracing
• Calculate the collective load balancing requirement based on the candidate nodes found by the agents
• Balance the cluster load

Practical applications: 

1. MASON library: java-based multi-agent simulation API library
2. Opt4J library: simple and intuitive GUI for loading meta-heuristic optimization models that can be applied for evolutionary algorithms

Resources: Artificial Intelligence for Big Data by Anand Deshpande and Manish Kumar