Simulation of evolution with artificial intelligence; An important step for the world of robotics

شبیه‌سازی تکامل با هوش مصنوعی؛ قدمی مهم برای دنیای رباتیک

A number of researchers are trying to find solutions that can simulate the gradual evolution of humans and other species of animals in natural environments in creatures based on artificial intelligence; But there are big and complex challenges ahead.

Hundreds of millions of years of evolution on Earth have created a vast diversity of different animal species. Each of the animal species on our planet has innate talents and skills, the ability to learn and acquire new skills, as well as specific physical characteristics. Different animals use all these innate and acquired traits to maintain their survival. Naturally, man is not separated from this rule.

But in the process of making creatures based on artificial intelligence, despite the fact that these creatures are inspired by nature and natural evolution, first the various elements related to intelligence and talent are created separately, and after all these elements are fully developed, together Are combined. Although such an approach has always had very good results; But it always limits the flexibility of creatures based on artificial intelligence to some of the simplest skills that even creatures with the simplest body structure can do well.

Researchers at Stanford University in the field of artificial intelligence have recently published a new article in the scientific journal Nature that offers a new technique that could help take a step forward in overcoming these obstacles. Researchers at the university claim that in this article, entitled Deep Evolutionary Reinforcement Learning, new techniques are proposed that first create a complex virtual environment and then the physical talents and abilities of creatures based on intelligence. Synthesis is strengthened and evolved. The results of this research can have important applications in the future in the field of research related to artificial intelligence and robotics.

Difficulty simulating the process of natural evolution

In nature, the brain and the body evolve simultaneously. The motor organs of each animal species have evolved over countless cycles and over several generations, and their nervous system manages and supports all the abilities that different species need to maintain and survive in their habitat. he does; Insects, for example, have a thermal vision system that detects the body temperature of other organisms.

In addition to having special wings, bats also have a natural echo system that helps them navigate in dark environments. Sea turtles also have plump legs that allow them to swim easily, and the animals’ magnetic field detection system helps guide them over long distances.

Man also has extraordinary abilities; Unlike many creatures, he can stand, and thanks to this innate feature, he can use his hands freely and see long distances easily. In addition, human fingers are delicate and fast, and human beings can use them to easily change flexible objects; Most importantly, man has a complex and completely unique brain compared to the brains of other creatures, which has made him the most social being on the planet, and he can find effective solutions to almost all problems.

It is interesting to note that all the animal species we see today are the evolved form of the first form of life that was created billions of years ago. Environmental stress in different environments has caused the first species of life to gradually evolve in different forms in order to acquire the necessary abilities to maintain their survival in their living environment. In a nutshell, the origin of all animal species is the first form of life created on planet Earth.

Studying and researching the evolution of talent and life can be very interesting and fascinating; But it is certainly very difficult to simulate such a process. If AI researchers want to intelligently simulate the life of different species, including humans, in exactly the same way that they evolved in nature, they must consider all possible forms of morphology, which is a very complex and computationally costly project. Becomes; Because doing so requires a lot of parallel and consecutive trial and error cycles.

Researchers in the field of artificial intelligence have identified shortcuts and pre-designed features to overcome such problems; For example, they first create a fixed architectural or physical design for an artificial intelligence or robotics system and then focus on optimizing trainable parameters. Benefiting from the school of evolution of Lamarckism instead of Darwinian theory of evolution is another shortcut method used by these researchers. Such an approach allows artificial intelligence-based creatures to transmit the abilities they have learned to later human beings.

Another approach used by researchers in the field of artificial intelligence is the separate training of artificial intelligence subsystems, which includes the visual, motor, linguistic and other systems. Researchers who use this approach will then integrate systems that have completed information training into the final system of artificial intelligence or robotics; Of course, although all of these approaches accelerate the process and reduce the cost of building creatures based on artificial intelligence with the ability to evolve gradually, but on the other hand, they also limit the flexibility and diversity of the results obtained.

Deep evolutionary reinforcement learning

Deep evolutionary reinforcement learning

Researchers at Stanford University plan to bring artificial intelligence research one step closer to the actual evolution process in a new research activity, while keeping costs as low as possible. They wrote in their article about this goal:

“Our goal is to clarify some of the principles governing the relationship between environmental complexity, evolved morphology, and the teachability of intelligence control.”

Their framework is called Deep Developmental Reinforcement (DERL). Human beings created within the framework of DERL based on artificial intelligence, use the deep reinforcement learning approach to acquire the skills needed to maximize their goals in life. In fact, researchers working on the basis of the DERL framework use the Darwinian school of evolution to find optimal solutions in various branches of morphology.

This means that when the next generations of creatures formed on the basis of artificial intelligence are created, the architectural and physical features taught to their predecessors are inherently transmitted to them with little change and mutation, and new generations do not need to learn them; Therefore, it can be said that the parameters taught to one generation of creatures based on artificial intelligence are not taught to the next generation, and these parameters become inherent features of new generations.

Researchers at Stanford University wrote in their article on the subject:

“DERL can provide the basis for large-scale in-computer experiments (experiments involving the simulation of biological systems in a computer). “These experiments can provide researchers with insights that help them understand how education and evolution work together to create complex relationships between the concepts of environmental complexity, morphological intelligence, and the teachability of control skills.”

Simulation of natural evolution process

Researchers at Stanford University use MuJoCo in their framework, which is in fact a virtual environment capable of providing highly accurate physical simulations without the possibility of deformation. The design space used by these researchers is called UNIversal aniMAL (UNIMAL), which aims to create different morphological forms to teach motor skills and skills related to changing and manipulating objects in different environments.

Each of the creatures created on the basis of artificial intelligence that is present in such an environment has a genotype (a series of genetic information) that defines the structure of their limbs and joints. The genotype of each of these organisms is passed on directly to the next generation in the form of innate characteristics. In the process of transmitting the genotype to the next generation, mutations and changes may occur in which major changes such as the creation of new limbs, the removal of some existing limbs or small changes in physical characteristics such as changes in the release of these organisms in motor activities and changes The size of the limbs follows them.

To test the results of the system they created, the researchers tested creatures based on artificial intelligence in three different environments, including the Flat (FT) environment, the Variable (VT) environment, and the Modifiable Objects environment. (MVT) becomes.

The first environment puts the least pressure on the creatures and they do not have to change their physical structure in this environment; But in the second environment, the situation is completely different and they have to change their physical structure to do things like climbing slopes or moving around obstacles in order to gain more mobility. In the third environment, in addition to strengthening their physical structure to overcome obstacles, they must also make changes to the objects in the environment to achieve their goal.

The benefits of deep developmental reinforcement learning

Deep evolutionary reinforcement learning
Deep evolutionary reinforcement learning in a variety of environments creates a variety of successful morphologies

One of the most interesting findings of DERL is its diversity. All the other approaches used to create an evolving artificial intelligence end in one solution; Because the new generation of creatures created on the basis of artificial intelligence directly inherit the physical features and skills learned by the previous generation; But in DERL, only morphological data is passed on to the next generation, and the end result of the system is the formation of a set of successful morphologies that include two-limbed, three-limbed, and four-limbed animals.

At the same time, there are signs of Baldwin effect in the system. Baldwin’s work is, in fact, an evolutionary theory that suggests that creatures that learn faster are more likely to inherently reproduce and transmit their characteristics to their next generation.

According to Stanford, DERL shows that only organisms are selected to evolve and improve, which learn these abilities without being directly pressured by the environment to learn different abilities.

Researchers at the university wrote in their article on the subject:

“It is interesting to note that Baldwin’s work could be used in the future to create new creatures with less complexity in selecting suitable specimens, as well as to create new generations with higher capacity.”

Deep evolutionary reinforcement learning
DERL-trained creatures are being evaluated to test their ability to perform various tasks from the top left, including patrolling, navigating to a specified point, exploring and searching, escaping, climbing a slope, pushing a box on a slope, and manipulating. Becomes a ball.

Finally, the DERL framework also examines the theory of intelligent trained organisms becoming more intelligent in a more complex environment. Stanford researchers have tested evolved creatures in eight different skills, including escaping, patrolling, changing objects, and exploring. The findings show that in general, creatures created on the basis of artificial intelligence trained in the three environments introduced in this paper have faster learning ability and better performance than creatures trained only in the flat environment.

These findings seem to be similar to the hypotheses put forward by DeepMind scientists. The AI ​​scientists believe that a sophisticated environment, an appropriate reward system, and reinforcing learning can ultimately lead to the emergence of all human-executable behaviors on the AI ​​platform.

Benefits of Deep Developmental Reinforcement Learning

The environment created based on the DERL framework has only a very small part of the complexity of the real environment.

Researchers at Stanford University wrote in their article on the subject:

“Although so far, by relying on DERL, we have been able to take a big step towards increasing the complexity of the environment used to increase the evolution of organisms based on artificial intelligence; But in the future, we need to design an environment with more open space that is more physically similar to the real world, and it is possible to experiment with multiple beings simultaneously. “Environmental design with these features is very important.”

The researchers plan to use more evaluation parameters in the future to see how AI creatures can enhance their ability to learn related human behaviors in a variety of contexts.

Research by Stanford scientists could have important applications in the future of artificial intelligence and robotics, and help researchers find ways to increase the evolution of AI creatures in ways that are more closely related to the natural evolution of organisms.

Researchers at Stanford University wrote in their article on the subject:

“We hope that what we have started will become an incentive to increase research and exploration in the field of developing the ability to learn and evolve in organisms based on artificial intelligence, and that this research will lead to more attitudes towards the ability of AI to learn human behavior quickly.” “And help increase progress in creating human behaviors in creatures based on artificial intelligence.”

More Similar Posts

Leave a Reply

Your email address will not be published. Required fields are marked *

Fill out this field
Fill out this field
Please enter a valid email address.
You need to agree with the terms to proceed

Most Viewed Posts
Techno barg We would like to show you notifications for the latest news and updates.
Allow Notifications