5 Ways mother nature inspires artificial intelligence
The world of AI has a lot of things around it to thank for its existence in our technological landscape of today. Not only have humans spent decades of research perfecting the mathematical calculations to make these wonderfully complex learning algorithms work but during this time we have looked further than our own species as inspiration to make the next generation of intelligent presence on our planet. Mother Nature, and all that it encompasses, has it’s roots firmly planted in the workings of Artificial Intelligence — and it’s here to stay.
Aren’t Sir David Attenborough’s wildlife documentaries just incredible? They go into incredible, high definition detail about the behaviours and properties of the Earth’s many inhabitants, and they allow us to understand how they fit into the natural ecosystem and work together in order to allow our planet to flourish — to make it Earth. Now I’m no Sir David Attenborough, but I’m still going to take you on a wildlife documentary of my own. The star creatures in question are none other than those Artificial Intelligence algorithms that are inspired directly by Mother Nature’s own. But first, I need to introduce you to two algorithm concepts. Search/Pathfinding and Predictive Modelling.
SEARCH (PATHFINDING) ALGORITHMS
Search algorithms are essentially programs that are designed to find the best/shortest route to an objective. For example, the travelling salesman problem is a typical search optimisation issue where you are given a list of cities and distances between those cities. You must search for the shortest route for the travelling salesman, whilst visiting each city once to minimise travel time and expenditure (ensuring that you return to the origin city). Real world applications of this problem are delivery trucks. Imagine 100 people in London made an online order and all the boxes are loaded into one van. The courier (let’s say…DPD), must now calculate the most efficient route (balancing distance/time taken) to deliver those packages from the depot (eventually returning to the depot) to ensure that the company is wasting as little time and money as it can during the delivery process.
PREDICTIVE MODELLING ALGORITHMS
Today, predictive modelling is where all the hype is. Data Scientists everywhere are shouting ‘Neural Networks!’ from the rooftops of their cosy office buildings and companies like Google are running around trying to solve the world’s problems with different variations of these complex little ‘artificial brains’. Essentially predictive modelling uses statistics in order to predict outcomes. You often hear Data Scientists attempting to solve two kinds of predictive modelling problems, Regression and Classification. Regression is the dark art of finding the correlation between two sets of variables, and classification is the process of determining the probability of a dataset belonging to a different set.
5 BIOLOGICALLY-INSPIRED LEARNING ALGORITHMS
1. ARTIFICIAL NEURAL NETWORKS
Algorithm Type: Predictive Modelling
Biological Inspiration: Cognitive Brain Function (Neurons)
Use-cases: Sentiment Analysis, Image Recognition/Detection, Language Correction, Robotics
Let’s start with the most commonly known Artificial Intelligence (AI) algorithm of them all. Neural Networks are part of a sub-category of AI called Machine Learning. They have been designed and built to mimic brain function on a Neuron level, interacting with Axons and Dendrites in order to pass information through a system to generate a predictive output through a series of ‘layers’. Each layer provides an extra layer of data representation and allows you to model most complex problems.
Neural Networks are probably the most widely used Machine Learning algorithms and is the hottest trend in Data Science and Machine Learning to date. The concept initially came to fruition in 1958, known as the Perceptron, later refined by Geoffrey Hinton and popularised by companies such as Google and Facebook. Neural Networks can be used to solve a wide range of problem types including Natural Language Processing, and Visual recognition. This supervised learning algorithm can support both Regression and Classification problems and examples of its use can be found in regular consumer products including smartphones, and connected home devices.
2. GENETIC ALGORITHMS
Algorithm Type: Search/Pathfinding
Biological Inspiration: Survival of the Fittest/Evolution (Cell Reproduction)
Use-cases: Data Mining/Analytics, Robotics, Manufacturing/Design, Process Optimisation
Genetic algorithms take a ‘survival of the fittest’, evolutionary-like approach among a series of individuals over consecutive generations in order to solve a search problem. Each generation contains a population of strings that mimic chromosomes that we see in our DNA. Each individual within a population represents a point within the search space, therefore each is a possible candidate solution. In order to improve the population of solutions, we put the individuals through a process of evolution.
- Each individual in the population will compete for resources and mates.
- The individuals most successful in each competition will (often) produce more individuals than those that perform poorly.
- Genes from more ‘ideal’ candidates propagate through the population so those good parents will often produce offspring with even greater potential.
3. SWARM/COLLECTIVE INTELLIGENCE
Algorithm Type: Search/Pathfinding
Biological Inspiration: Ant Colonies/Schools of Fish/Flocks of Birds
Use-cases: Robotics, Video Game AI, Manufacturing, Route Planning
Ant Colony Optimisation and Particle Swarm Optimisation are the two most commonly known algorithms that conform to the bounds of the ‘collective intelligence’ concept. On a basic level, the algorithms in question all utilise multiple working agents. Each working agent exhibits very basic behaviour that collectively (as a pack) works together to bring out more complex, emergent behaviour in order to solve the problem.
Ant Colony Optimisation (ACO) is very different to Particle Swarm Optimisation (PSO). Both aim to achieve emergent behaviour but go about it in two different ways. Like real ant colonies, ACO utilises a pheromone scent to guide individual agents onto the shortest path. Initially, a random pheromone is initialised across the problem space. Individual agents will begin to traverse the search space, dropping pheromone scent as they go. Throughout each time-step, the pheromone will decay at a defined rate. Single Agents make their decisions based on the strength of pheromone scent in front of them in order to traverse the search space. The stronger the scent in a particular direction, the more likely it is to travel that way. The best-known solution will end up being the one with the strongest pheromone scent.
PSO looks more to the overall direction of the pack. A number of single agents are initialised and they begin in random directions. At each time step, each agent needs to make a decision on whether or not to change direction. The decision will be based on the direction of the best-known solution (known as pbest/global best), the direction of the best nearest neighbour (local best), and the current direction of travel. The new direction of travel is typically a good ‘compromise’ of all these values.
4. REINFORCEMENT LEARNING
Algorithm Type: Predictive Modelling
Biological Inspiration: Classical Conditioning
Use-cases: Video Games, Controlling Autonomous Vehicles, Manufacturing Line Software, Financial Systems
With groundings in psychology and a process reminiscent of classical conditioning, Reinforcement Learning (RL) supports the idea of providing a positive numerical response to a favourable action taken by an agent. Learning the concept of Reinforcement Learning is often easier when compared to the popular classical conditioning example; Pavlov’s Dogs. It was a piece of research conducted in the 1890’s where Russian Psychologist Ivan Pavlov was looking at the salivation in dogs in response to being fed. An excellent article explaining the matter can be found here. Essentially, if an RL agent takes a good action, one that makes a step towards completing the task required, it is given a numerical reward. The agent will learn using a policy — for example; to maximise the reward given at each individual step. The application of raw inputs into the algorithm allows the agent to develop its own sense of the problem, and how to solve it in the most efficient manner.
It is common to see RL algorithms paired with other Machine Learning technologies — such as Neural Networks. This is often referred to as Deep Reinforcement Learning. Neural Networks are often used to estimate the reward given to an RL agent when it makes a particular decision. Deep Mind, now a Google company, has made great progress within this space, using a Deep Q Learning approach to tacking more generic problems (such as the ability for one algorithm to play an entire library of Atari games, unassisted and beat the World Champion itself at the game ‘Go’. They are now adapting this approach to be able to tackle more complex games such as Starcraft II.
For reference, Q Learning is a model-free version of the Reinforcement Learning algorithm. It can be used to find the optimal action-selection policy for any finite Markov Decision Process. Upon program initialisation, the Q values for each action-value pair are defined by the developer and updated by the RL algorithm at each time step. The following image is an example of the equation to update the Q action-value pair.
5. ARTIFICIAL IMMUNE SYSTEMS
Algorithm Type: Predictive Modelling
Biological Inspiration: Immune Systems
Use-cases: Security Software, Autonomous Navigation Systems, Scheduling Systems, Fault Detection Software
An immune system is a system that protects the body against substances and pathogenic organisms by producing an immune response. Artificial Immune Systems (AIS) are adaptive systems, inspired by theoretical immunology and observed immune functions that are applied to problem-solving. AIS is a subfield of biologically-inspired computing and natural computation with links to Machine Learning and Artificial Intelligence. There are multiple algorithms commonly associated with AIS:
- Clonal Selection
- Dendritic Cell
- Negative Selection
- Artificial Immune Recognition
Like the biological immune system, AIS is able to categorise all ‘cell’s within a system as ‘self’ or ‘non-self’ cells. A distributed task force of intelligence is used to take action upon all cells. The two most important types of cells involved in immunisation are B-cells and T-cells (white blood cells to you and me). T-cells come in three types; one type to activate B-cells, one type to bind to and destroy foreign invaders and a final type to suppress auto-immune issues. B-cells are responsible for producing antibodies which are specific proteins that bind to an antigen — a toxic/foreign substance. Artificial Immune Systems are often used to defend against cyber attacks by monitoring intrusion detection and are often integrated into enterprise-grade software. Unlike the other algorithms mentioned in this article, free online learning material on this subject is very limited and is likely the least-developed.
LET’S WRAP UP…
So I’ve given 5 examples of how Mother Nature inspires our research and enhances our technology. There are many more biologically-inspired algorithms that influence our AI systems, so please share your experiences and knowledge of those I’ve missed in the comments section.
I am a Fitness Athlete & Software Engineer with a passion for Bodybuilding and Machine Learning. I run my own blog and website where I discuss these topics — and more. I provide links to my Github repository so you can see my work-in-progress projects and social media accounts (Facebook, Instagram, Twitter etc.) so you can follow my fitness transformation.
http://lukealexanderjames.com/
If you like what you read, please consider holding down the clap button for a while!
