Models of artificial life. Part 1
Immediately to the point. What can the study of IL ? How to practically apply the knowledge gained in this area? Scientists working in this area say that their research is aimed specifically at practical applications, such as mobile robots, medicine, nanotechnology, artificial intelligence, and even research into social and economic systems. But in the history of artificial life there were quite a few disputes and contradictions. Many skeptics criticize some of the work on artificial life, calling them "science without facts." However, recent publications on artificial life suggest that the technologies used to model artificial life are recognized by the scientific community, at least to study evolution.
The main problem of IL is the creation of computing systems and models operating on the basis of biological and evolutionary principles. Simply put, it is “the world created by man.” But until a certain time, work in this area was limited only to theoretical studies of the properties of life. And modeling is only the construction of individual parts of biological systems, such as, for example, adaptive behavior or evolution . But recently, largely due to the development of nanotechnology and molecular biology, as well as due to the increased understanding of the purpose of individual genes and how they interact, and the emergence of means of manipulating individual molecules, a new direction of IL, aimed at creating new, artificially synthesized biological forms, is being developed, which requires the philosophical substantiation of the IL and, at a minimum, the definitions of the term “life” (what is generally understood by this word). Today, supporters of this trend say that they are exploring more general forms of life than those that exist on Earth. Let's try to track the approximate course of development of IL - where it all began and what this trend may lead in the future.
So, it became clear to us that artificial life is nothing more than an attempt to study life itself, living systems, their evolution, with the help of man-made programs. But it is interesting how all this can work in practice: in medicine, robotics, to study the “life” of social and economic systems, etc.
To begin with, let us consider one of the first models of artificial life, in which the software modeling of fairly simple processes of organisms was carried out - this is PoliMir Larry Yager.
The model can be imagined as a large table bounded by walls at the edges. On this table live artificial organisms. They move, absorb food, interbreed, giving offspring. Also, organisms can come to grips with others, while the defeated turns into food. Yager assigned organisms an interesting feature - they all have color and color vision. For example, if an organism enters into a struggle, then it becomes red, and if it experiences a desire to cross, it becomes blue.
All these properties allow the population of these organisms to evolve, which helps us, in fact, to feel, see and analyze similar evolutionary processes. In fact, the Jager model showed:
Summing up the small summary of the previous chapter, we can say that Larry Yager, with the help of his PoliMir, realized nothing more than the natural selection of artificial organisms and nothing more. This is what is meant by the word "Evolution" in PoliMir. Ackley and Littman tried to extend the term a bit. Their work is a model of interaction between learning and evolution.
The general pattern of agent behavior is quite simple. Agents live in a two-dimensional world, divided into cells. Agents, trees, predators, grass and stones can be located in cages. Accordingly, agents are the main “unit” of the world, and everything else influences them in various ways.
Predators beat agents, and more than agents of predators. The means for protection from predators is a tree, on which an agent can climb if there is no other agent there. But it is also dangerous to be constantly on a tree, as trees eventually die, killing the agent sitting on it.
Each agent has its own neural network, with the help of which it makes decisions about its actions. The agent's neural network consists of two blocks:
Actually, these blocks and allow you to combine training and evolution in this model. The action assessment unit “generates” a training signal for the behavior unit, which in turn causes the agent to act accordingly. The signal itself can be either positive, if the evaluation unit “believes” that the team of the behavior unit has improved the life of the agent, or negative - otherwise. Thus, after various successes and failures of the agent, his neural network is corrected and takes the “right” decisions in the same actions. This is how learning is done.
And in the process of crossing two agents, their neural networks are transmitted to descendants - this is how the evolution in the model is carried out.
The results of Ackley and Littman were very, very interesting. They started the model in turn: first, the model was only with training, then the model was only with evolution, and finally the model was complete, with both training and evolution. It became clear that with the interaction of training and evolution, agents did not die out during a million life cycles (and once the number of cycles reached 9 ∙ 10 6 ), while with single models, agents died out 5-7 times faster!
')
PS I would like to add a review of such models as "Grasshopper", "Ant Farm" and "Tierra". Habravchanam interesting?
PPS Probably, it makes sense to transfer the post to the blog Artificial Intelligence , but I did not succeed, t.ch. I beg your pardon.
UPD Thank you all so much for the karma that allowed us to publish an article in the right blog and for comments! I start writing part 2.
UPD If there is no desire to collect PoliMir, then you can just watch the video and graphics .
The main problem of IL is the creation of computing systems and models operating on the basis of biological and evolutionary principles. Simply put, it is “the world created by man.” But until a certain time, work in this area was limited only to theoretical studies of the properties of life. And modeling is only the construction of individual parts of biological systems, such as, for example, adaptive behavior or evolution . But recently, largely due to the development of nanotechnology and molecular biology, as well as due to the increased understanding of the purpose of individual genes and how they interact, and the emergence of means of manipulating individual molecules, a new direction of IL, aimed at creating new, artificially synthesized biological forms, is being developed, which requires the philosophical substantiation of the IL and, at a minimum, the definitions of the term “life” (what is generally understood by this word). Today, supporters of this trend say that they are exploring more general forms of life than those that exist on Earth. Let's try to track the approximate course of development of IL - where it all began and what this trend may lead in the future.
PolyWorld
Posted by: Larry Yaeger, 1993
So, it became clear to us that artificial life is nothing more than an attempt to study life itself, living systems, their evolution, with the help of man-made programs. But it is interesting how all this can work in practice: in medicine, robotics, to study the “life” of social and economic systems, etc.
To begin with, let us consider one of the first models of artificial life, in which the software modeling of fairly simple processes of organisms was carried out - this is PoliMir Larry Yager.
The model can be imagined as a large table bounded by walls at the edges. On this table live artificial organisms. They move, absorb food, interbreed, giving offspring. Also, organisms can come to grips with others, while the defeated turns into food. Yager assigned organisms an interesting feature - they all have color and color vision. For example, if an organism enters into a struggle, then it becomes red, and if it experiences a desire to cross, it becomes blue.
All these properties allow the population of these organisms to evolve, which helps us, in fact, to feel, see and analyze similar evolutionary processes. In fact, the Jager model showed:
- Difficult color vision could well have been formed in the process of evolution.
- Perhaps the evolutionary formation of three strategies: "stupid cow", "lazy cannibal" and "life on the edge of the world."
« » — , , , .
« » — , , , ( ).
« » — , , , , , , .Interaction between learning and evolution
Authors: D. Ackley, M. Littman, 1992
Summing up the small summary of the previous chapter, we can say that Larry Yager, with the help of his PoliMir, realized nothing more than the natural selection of artificial organisms and nothing more. This is what is meant by the word "Evolution" in PoliMir. Ackley and Littman tried to extend the term a bit. Their work is a model of interaction between learning and evolution.
The general pattern of agent behavior is quite simple. Agents live in a two-dimensional world, divided into cells. Agents, trees, predators, grass and stones can be located in cages. Accordingly, agents are the main “unit” of the world, and everything else influences them in various ways.
Predators beat agents, and more than agents of predators. The means for protection from predators is a tree, on which an agent can climb if there is no other agent there. But it is also dangerous to be constantly on a tree, as trees eventually die, killing the agent sitting on it.
Each agent has its own neural network, with the help of which it makes decisions about its actions. The agent's neural network consists of two blocks:
- The block of behavior, i.e. a block defining the agent's actions at each tick of time.
- Action evaluation unit, i.e. block that forms the goals of the agent's behavior.
Actually, these blocks and allow you to combine training and evolution in this model. The action assessment unit “generates” a training signal for the behavior unit, which in turn causes the agent to act accordingly. The signal itself can be either positive, if the evaluation unit “believes” that the team of the behavior unit has improved the life of the agent, or negative - otherwise. Thus, after various successes and failures of the agent, his neural network is corrected and takes the “right” decisions in the same actions. This is how learning is done.
And in the process of crossing two agents, their neural networks are transmitted to descendants - this is how the evolution in the model is carried out.
The results of Ackley and Littman were very, very interesting. They started the model in turn: first, the model was only with training, then the model was only with evolution, and finally the model was complete, with both training and evolution. It became clear that with the interaction of training and evolution, agents did not die out during a million life cycles (and once the number of cycles reached 9 ∙ 10 6 ), while with single models, agents died out 5-7 times faster!
')
PS I would like to add a review of such models as "Grasshopper", "Ant Farm" and "Tierra". Habravchanam interesting?
UPD Thank you all so much for the karma that allowed us to publish an article in the right blog and for comments! I start writing part 2.
UPD If there is no desire to collect PoliMir, then you can just watch the video and graphics .
- Part 1 (PolyWorld, Interaction between learning and Evolution)
- Part 2 (Grasshopper, AntFarm)
- Part 3 - coming soon ...
Source: https://habr.com/ru/post/55119/
All Articles