Source: French to English Tester Published on: 2026-05-13
Source: The Conversation – France in French (2)– By Karol Desnos, Senior Lecturer at the Institute of Electronics and Digital Technologies, INSA Rennes AI systems are often criticized for their high energy consumption — to the point that one wonders what electrical production will power the data centers planned worldwide.
But more energy-efficient artificial intelligence systems do exist. Some are inspired by nature's evolutionary strategy to arrive at the simplest possible solution for a given problem. The principles of natural selection theorized by Charles Darwin have enabledthe evolution of living beings remarkably adapted to their environments.
The human brain is a remarkable example of the outcome of this evolution, consuming only about twenty watts,which is nearly 1000 times less than an artificial intelligence surpassing it at the game of Go. By digitally reproducing these evolutionary principles,it is possible to build more efficient artificial intelligences (AI).
The rapid advances of AIs during this past decade are mainly due to the use of artificial neural networks, called "deep" networks, capable of effectively learning a wide variety of tasks.
However, executing such a network requires performing several million to several billion mathematical operations by a computer; and the greater the complexity of a task, the larger the size of the network needed to accomplish it will be as well.
The massive use of deep neural networks, increasingly large to enhance their capabilities, poses a problem ofmajor sustainability. Indeed, these networks require increasing natural resources and energy: processors executing these calculations must be produced, powered, and cooled, notably theGPU.
To meet this sustainability challenge, it is necessary to propose alternative AI technologies that naturally adapt their complexity to that of the task performed in order to be more efficient. Evolutionary approaches for AI learning offer a credible alternative to deep neural networks, being more efficient while remaining effective.
Indeed, unlike neural networks whose size is fixed by a developer before training, the evolutionary approach builds an AI whose size minimally adapts to a specific task. The result is an AI with a computational complexity several hundred to thousand times lower than neural networks, and thus naturally more efficient.
But by the way, what exactly is evolution? Natural selection and evolution rely on three essential elements: individuals defined by a genome, a reproduction mechanism, and a selection process. In biology, the genome of each living being is constructed by assembling basic building blocks common to a wide variety of species: DNA.
The genome of a being characterizes it as an individual, determining many of its traits: shape, physiology, size, colors; and predisposes it to certain behaviors, such as running or swimming. The reproduction mechanism allows one or more individuals to give birth to new individuals by copying and mixing their genomes.
This copy, sometimes imperfect, creates a new individual who possesses its own genome, conferring traits resembling those of its parents but possessing its own characteristics. A group of individuals forms a population, which exists in an environment where they are constantly tested through challenges: searching for nutrients, surviving in a hostile environment, finding a reproductive partner.
Some traits will make individuals more successful in these challenges, increasing their chances of survival and reproduction, while other less favorable traits tend to disappear: this is natural selection. Generation after generation, this long process has led to the emergence on Earth of living beings adapted to their environments.
For nearly 60 years, thescientific study of evolutionary algorithmsaims to take up these principles of evolution for the optimization of systems or artificial objects.
A concrete use case example isthe optimization of airplane wings, where the "genome" characterizes the profile, length, and width of a wing; and where the evaluation of an individual (a wing model, in this case) measures its resistance, lift, and weight.
How to apply evolution principles to build AIs Genetic programming is a scientific field aimed at building computer programs, including AIs, by applying these evolutionary principles. Image Credit: Nicolas Beuve In its simplest form, the "DNA" used to create an individual is a set of instructions or basic mathematical functions: addition, multiplication, cosine, etc.
The genome of each individual is thus a sequence of instructions, called a program, which performs calculations on environmental data. Let's imagine, for example, that we want to build an AI in charge of controlling a robot.
The AI observes numbers representing the current position of the different parts of the robot, the angle of its joints, and the speed of these different elements. These numbers are used to execute the AI's individual program.
The results of the latest instructions constitute the AI's response to this observation and are used to control the robot's various motors. The evolutionary process begins with the creation of a population of individuals by generating short random programs, a simple addition for example.
The selection of the best programs is done by keeping those most capable of performing the desired task, for example, moving a robot as far as possible. In the early generations, even the best individuals are generally very poor, but they form the genetic capital for the first reproduction phase.
A program can reproduce by crossover, interleaving instructions from two parent programs; or by mutation, reproducing an existing program imperfectly to add or remove an instruction. This process is repeated over many generations, and at the end of the evolution process, the program of the best individual is preserved to be used as AI.
So, are the AIs obtained through this process of evolution more frugal? During the evolutionary process, the number of instructions, and thus the complexity of programs, automatically adapts to the difficulty of the task to be performed.
Indeed, the addition of new instructions to individuals' genomes persists only if it provides them with better abilities, promoting their survival and reproduction. Thus, the evolutionary process naturally favors the emergence of programs with few instructions, yet well adapted to the task.
The use of programs constructed in this way does not require a dedicated GPU-type chip, and can generally be performed on existing low-power processors, and therefore more energy-efficient ones.
Another advantage of the evolutionary approach: "interpretability" While traditional neural networks are capable of performing complex tasks, their high computational complexity often makes it impossible tointerpretthe causes of their proper functioning, or worse, of their errors.
Again, the brevity of programs resulting from the evolutionary process is a major asset, since this makes it possiblethe clear interpretation of the functioning of the AI thus created. An evolutionary algorithm is by nature easy to interpret.
Source: Quentin Vacher.
This video presents an example, with a program created to illustrate this article that allows controlling a robotic leg called the "hopper." Usually learned with complex neural networks, genetic evolution enabled an AI to learn to move the robotic leg using very simple instructions to control each of the robot's three motors.
Indeed, understanding the causality of the AI's actions based on observations is possible, and we see that the program demonstrates great logic where each motor is mainly controlled by limbs close to it. What future for AIs resulting from evolutionary processes?
In certain application domains, more efficient AIs resulting from an evolutionary processcompete with the capabilities of neural networks for one hundredth (or even one thousandth) of their cost, for example inroboticsor inthe cyber defense industry.
If the cost of these AIs resulting from the evolutionary process makes them intrinsically more frugal, it is neverthelessensure that this sobriety does not lead to a rebound effect, in the form of an even more massive use of such AI for applications where it is not strictly necessary.
This research field offers many opportunities and challenges to the scientific community, whose small size cannot compete with the colossal investments around neural networks.
Among these challenges, scaling up evolutionary process-based AIs that still fail to compete with neural networks on the most complex tasks, such ascontrol of humanoid robots, or thenatural language processing at the core of conversational bots.
The projectfouticsis supported by the National Research Agency (ANR), which funds project-based research in France. The ANR's mission is to support and promote the development of fundamental and applied research in all disciplines, and to strengthen the dialogue between science and society.
For more information, visit the website of theANR. < class="fine-print">Karol Desnos received funding from the National Research Agency (ANR) under project ANR-22-CE25-0005-01. < class="fine-print">Mickaël Dardaillon received funding from the National Research Agency (ANR) under the project ANR-22-CE25-0005-01. < class="fine-print">Quentin Vacher received funding from the National Research Agency (ANR) under the project ANR-22-CE25-0005-01. –ref.
Evolutionary algorithms, a way to make AIs more efficient –https://theconversation.com/evolutionary-algorithms-a-path-to-make-ais-more-sober-281088