AI is now learning evolution as terrestrial lifestyles

AI is now learning evolution as terrestrial lifestyles.jpgsignature169821ae58721fe142dc914eb528906f

This article is part of our reviews of AI research papers, a series of posts that examine the latest findings in artificial intelligence.

Hundreds of millions of years of evolution have blessed our planet with a wide variety of life forms, each discerning in its own fashion. Each species has evolved to develop indigenous skills, learning abilities, and physical form that will ensure its survival in the environment.

But despite being inspired by nature and evolution, the field of artificial intelligence has largely focused on creating the elements of information separately and linking them together after development. . While this approach has yielded good results, it has also limited the flexibility of AI reps in some of the basic skills found in even the simplest of lifestyles.

In a new paper published in the scientific journal Nature, AI researchers at Stanford University present a new invention that will help take steps toward overcoming some of these limitations. Entitled “Deep Evolutionary Reinforcement Learning,” the new approach uses a complex virtual environment and reinforcement learning to create meaningful representations that can evolve both physically and physically. learning abilities. The decisions for the future of AI research and robotics can have a major impact.


    It's hard to imitate

    Credit: Ben Dickson / TechTalks

    In nature, the body and the brain grow together. Over many generations, every animal species has undergone innumerable mutations to grow organs, organs, and the nervous system to support the functions it needs in its environment. Mosquitoes have a thermal view to detect body heat. Bats have wings for flight and an echo device to navigate dark places. Sea turtles have flippers for swimming and a magnetic field detection system for long distance travel. Humans have an upright position that frees their arms and allows them the longest horizons, flexible hands and fingers that can handle objects, and brains that make them the best social creatures and rescuers. problems on the planet.

    Interestingly, all of these species came from the first life-form that appeared on Earth several billion years ago. Based on the selective pressures exerted by the environment, the descendants of these first living creatures evolved in many ways.

    It is interesting to study the evolution of life and intelligence. But reproduction is very difficult. An AI system that would want to recreate an intelligent life in the same way as evolution would have to find a vast array of possible morphologies, which are computationally expensive. It required a lot of parallel trial-and-error rounds.

    AI researchers use a number of preloaded snapshots and features to overcome some of these challenges. For example, they repair the architecture or physical design of an AI or robotic system and focus on making good use of the learnable parameters. Another shorthand is the use of Lamarckian rather than Darwinian evolution, in which AI representatives pass on their learned parameters to their descendants. But another way is to train different AI subsystems (vision, locomotion, language, etc.) and then proceed together in a final AI or robotic system. While these approaches will speed up the process and reduce training costs and the evolution of AI reps, they will also limit flexibility and a mix of achievable results.

    READ  Microsoft Viva offers modules for learning, information sharing, and more

    Deep reinforcement learning of evolution

    Reinforce deep learning structure
    Credit: Ben Dickson / TechTalks

    In their new work, the researchers at Stanford aim to take AI research one step closer to the evolutionary process while at the same time keeping costs as low as possible. “Our aim is to explain some of the principles that govern the relationship between environmental complexity, morphology of change, and controllable learning ability,” they write in their paper.

    Their framework is called Deep Evolutionary Reinforcement Learning. In DERL each agent uses in-depth reinforcement learning to develop the skills needed to further his or her goals throughout his or her life. DERL uses Darwinian evolution to find the morphological space for the best solutions, which means that when a new generation of AI agents spawn, they possess only the physical and architectural features of the parents (along with minor mutations). None of the learned parameters are passed down over generations.

    “DERL opens the door to major experiments in silico experiments to provide scientific insights into how learning and evolution create solemn relationships between environmental complexity, morphological knowledge, and ability learning control activities, ”the researchers write.

    Simulating evolution

    For their framework, the researchers used MuJoCo, a virtual environment that provides a highly accurate solid-state physics simulation. Their design space is called UNIversal aniMAL (UNIMAL), in which it aims to create a morphology that studies locomotion functions and object handling in a number of areas.

    Each agent in the environment is made up of a genotype that defines its boundaries and joints. The direct descendants of each agent inherit a parent's genotype and go through spells that can create new members, remove existing members, or make minor changes to features such as freedom levels or organ size.

    Each agent is trained with reinforcement learning to get the most benefits in different environments. The most basic work is a locomotion, in which the agent is rewarded for the speed at which he or she travels in a program. Representatives with a more physically structured structure for crossing terrain learn more quickly to use their arms to move around.

    To test the results of the system, the researchers generated representatives in three types of ground: horizontal (FT), variable (VT), and variable ground with variable (MVT) terrain. The flat ground places the least selective pressure on the morphology of the producers. The variable fields, on the other hand, force the producers to develop a more complex physical structure that can climb slopes and move around obstacles. The MVT difference has the added challenge of requiring the producers to manipulate items in order to achieve their goals.

    Benefits of DERL

    Benefits of DERL
    Credit: Ben Dickson / TechTalks

    We use cookies to ensure that we give the best user experience on our website. If you continue to use this site we will assume that you agree. More information