Throughout the history, intelligence was defined in many ways: the ability to do math, to solve problems, understanding communication, self-awareness, the ability to learn, and so on, at first in the effort to establish the difference between humans and animals (intelligence being one of the main differences), and later in order to understand life as a whole.
Our current understanding of intelligence can be described as the ability to perceive information, to retain knowledge and even apply it for behavior adaptation (https://en.wikipedia.org/wiki/Intelligence). In this sense, all known life is intelligent, as adaptations to available environments meant changing behavior in order to survive.
Artificial Intelligence (or AI), on the other hand, has vastly different applications. Throughout the still short history of the IT world, computer scientists have started to develop algorithms for hard mathematical problems, for which an established algorithm has not been designed yet. Thus, generalized approaches have come to the rescue.
Types of problems
In math, we have three elements that define a completely solved problem: the input, the objective function or some way of computing output based on the input values, and the output itself (see image below).
In the AI world, we have two types of problems, which need to be solved in some way:
- Problems for which we know the outputs and the inputs but we don’t know how to calculate the outputs based on the inputs (regression class problems, classification and so on).
- Problems for which we know the inputs and the objective function (a way to estimate the fitness of a proposed solution), and we want to optimize the input parameters in order to get the best outputs (TSP/VRP class and many other).
Taking inspiration from nature, there are two main algorithmic approaches, one for each type of problem:
- Evolutionary algorithms, that use concepts like evolution but also animal-world-approaches like the behaviour of ant colonies or bees in order to optimize different functions.
- Machine Learning approaches, that "learn" the data and come up with a way of computing the outputs based on the input.
Evolutionary Optimization, Genetic Algorithms
From the optimizations world, one of the best known class of algorithms are the evolutionary ones, and especially the Genetic Algorithms (GA) class. GA's use the principles of evolution, formulated by Darwin (but also with the latest observations made by biology experts), in order to generate (at least) a viable solution and then optimize it (improve a certain performance indicator). We can apply such algorithms in problems where we have the inputs, we have a way of knowing if a solution is good or not, but we do not have a way of computing the best solution.
For example, let's study a Travelling Salesman Problem (TSP). A salesman wants to visit a bunch of cities, while keeping the costs as low as possible. In order for this to happen, the salesman needs a route. A route can be defined as the order in which the cities must be visited. The best route is this order of cities that results in the smallest cost, where the cost can be time, money, distance, whatever we are interested in.
In a GA implementation, each possible (or candidate) route is called an individual. Each city is a Gene inside the individual's Chromosome, the order of the genes representing the final route. The algorithm works by creating a population of individuals and applying a set of operators: Selection, Crossover and Mutation. Recently, other aspects are being considered as operators, like: Survival, Inbreeding, Monogamy, etc.
As intuition might tell you, evolutionary algorithms work by gradually changing parts of individuals in a pool of solutions, in order to see if the results are better. Thus, we do not have a way of constructing the best route, but we do have a way of knowing if the route is good (by summing the costs between cities).
Machine Learning Approaches
In contrast with optimization problems, machine learning approaches a vastly different problem area. For instance we might want to teach software how to read (Optical Character Recognition, a classic clasification problem), or create a piece of software that can filter spam while also evolving together with the spammer's creativity, or even use historic data in order to predict the traffic congestion that might happen tomorow at noon.
In order for this to happen, the software piece has to guess the relationship between the inputs and the output. One of many approaches is to use Artifficial Neural Networks (ANN), that are a basic copy at the behaviour level, of the human brain. Neurons and synapses (connections between neurons) are simulated in an algorithmic fashion and simulate the way the brain learns. Learning is called training and this prepares the ANN for the specific application that the engineer builds.
The biggest advantage of this approach is the ML model's ability to generalize. This means that if you train it on some data that relevantly describes a problem, the ML model will be able to apply the same knowledge for other data that will be presented to its inputs (of course, with a certain accuracy).
Artifficial Intelligence is more than just computers understanding what we say to them, or performing tasks intelligently. There are hard problems for which we do not yet have specialized algorithms, or, if we do have, those algorithms are very slow (like Backtracking, a Brute Force approach). In some cases, AI approaches can speed up problem solving, or provide ways to solve them, no matter how complex the problem is. Of course, good engineers are a must, as designing and implementing such systems require great care and intensive knowledge.
Holisun can either help designing complete systems or offer assistance into algorithm and system design for all these kinds of algorithms. The knwoledge landscape that the Holisun team has gathered throughout the years encompases both kinds of AI approaches: Evolutionary Optimizations as well as Machine Learning approaches. Hardware to AI interactions are also fields of study for us, we have multiple projects that tackle this subject, like our MUSNOMICS Research Project, or GOHYDRO Research Project.