Models of artificial life. Part 2
At the request of the Habrachians continued part 1
It is no secret to anyone that purposeful behavior is inherent in the entire animal world, including man, i.e. the desire to achieve certain goals. For animals, this is most often survival and reproduction. In their work, Burtsev, Gusarev and Redko try using a computer model to answer the question: "How, in principle, purposeful behavior could arise in the process of evolution of life on our planet?"
In this project, the authors use such a thing as motivation . Indeed, the motivation of any living creature that falls into a particular situation stimulates the “right” decision-making. Using motivation, the model explores the possible mechanism for the occurrence of purposeful behavior in the process of evolution.
All actions in the model unfold in a one-dimensional cellular environment. Time is discrete - i.e. one action is performed at each time step. In cells, with a certain probability, grass grows. There is a population of agents that have a need for energy, due to nutrition and the need for reproduction. The agent’s energy is spent in the following actions, and different amounts of energy are spent on different actions:
Each agent's need is characterized by quantitative motivation. For example, if an agent sees another agent nearby and its energy resource is sufficient for reproduction, it marks itself as ready for reproduction, if the second agent does the same, a cross occurs. As a result, a new agent appears that takes parts of the energy resource from the parents. Each agent has its own neural network, which has special inputs from motivation. Due to the neural network, the behavior of the agent is controlled and the evolution of agents also takes place - the genome (a set of neural network weights) of the offspring is formed, as a result of crossing, on the basis of the parents genome by recombination and mutation. When the resource is reduced to zero, the agent dies. For analysis, two options were modeled: agents with motivation and agents without motivation. Different parameters P were also set - the probability of an accidental appearance of grass at each time. The final results showed that the population of agents with motivation is much better adapted to the environment, and with an average amount of food (P = 1/200), the population of agents with motivation “finds” a fairly effective survival strategy, and the population without motivation dies out completely.
The results are logical and very understandable, because for an agent without motivation, all that remains is that at the sight of food to eat it, at the sight of a neighbor - cross with him, and in the absence of everything, stand and do nothing (rest), which leads the agents to imminent death, with a small or medium amount of food. When the agent has motivation, in the process of evolution, the agent begins to act in approximately the following way: there is not enough resource - to look for food or rest, a lot of resource - to perform any actions. Due to this scheme, the population survives much more efficiently than in the case without motivation.
Many, observing the behavior of ants, pay attention to the work of ants. When they get food, they take it to the anthill, where it is processed and eaten by all members of the colony. Many species of ants exhibit a high degree of coordination and cooperation between “prey” (usually through pheromone transfer of information).
The "ant farm" mimics the evolution of complex behavior in artificial organisms. It considers organisms that live and reproduce in a relatively complex environment, with many senses (internal and external), and a large set of possible actions at any given time. Moreover, each ant has a memory, and, therefore, its behavior depends on its history. Throughout life, each ant is born, manages to make thousands of different decisions and actions, leaves offspring (this depends on the behavior of the organism throughout its life), and eventually dies.
By the way, the AntFarm model was developed in CM ++ based on the Connection Machine 2 supercomputer , which specializes in developments in the field of Artificial Intelligence.
UPD Article, in English. It more or less in detail describes the model of Collins and Jefferson - AntFarm.
Grasshopper
Authors: M. S. Burtsev, R. V. Gusarev, V. G. Redko, 2002
It is no secret to anyone that purposeful behavior is inherent in the entire animal world, including man, i.e. the desire to achieve certain goals. For animals, this is most often survival and reproduction. In their work, Burtsev, Gusarev and Redko try using a computer model to answer the question: "How, in principle, purposeful behavior could arise in the process of evolution of life on our planet?"
In this project, the authors use such a thing as motivation . Indeed, the motivation of any living creature that falls into a particular situation stimulates the “right” decision-making. Using motivation, the model explores the possible mechanism for the occurrence of purposeful behavior in the process of evolution.
All actions in the model unfold in a one-dimensional cellular environment. Time is discrete - i.e. one action is performed at each time step. In cells, with a certain probability, grass grows. There is a population of agents that have a need for energy, due to nutrition and the need for reproduction. The agent’s energy is spent in the following actions, and different amounts of energy are spent on different actions:
- To be at rest (to rest) is the least amount of energy;
- Eat (replenish your energy resource) - twice as much energy;
- Move i.e. move one cell to the right or left - two times more energy costs than food;
- Jump through several cells in a random direction - 5 times more cost than when moving;
- Crossing is also 5 times more energy than when moving.
Each agent's need is characterized by quantitative motivation. For example, if an agent sees another agent nearby and its energy resource is sufficient for reproduction, it marks itself as ready for reproduction, if the second agent does the same, a cross occurs. As a result, a new agent appears that takes parts of the energy resource from the parents. Each agent has its own neural network, which has special inputs from motivation. Due to the neural network, the behavior of the agent is controlled and the evolution of agents also takes place - the genome (a set of neural network weights) of the offspring is formed, as a result of crossing, on the basis of the parents genome by recombination and mutation. When the resource is reduced to zero, the agent dies. For analysis, two options were modeled: agents with motivation and agents without motivation. Different parameters P were also set - the probability of an accidental appearance of grass at each time. The final results showed that the population of agents with motivation is much better adapted to the environment, and with an average amount of food (P = 1/200), the population of agents with motivation “finds” a fairly effective survival strategy, and the population without motivation dies out completely.
The results are logical and very understandable, because for an agent without motivation, all that remains is that at the sight of food to eat it, at the sight of a neighbor - cross with him, and in the absence of everything, stand and do nothing (rest), which leads the agents to imminent death, with a small or medium amount of food. When the agent has motivation, in the process of evolution, the agent begins to act in approximately the following way: there is not enough resource - to look for food or rest, a lot of resource - to perform any actions. Due to this scheme, the population survives much more efficiently than in the case without motivation.
Antfarm
Authors: Robert J. Collins, David R. Jefferson, 1991
Many, observing the behavior of ants, pay attention to the work of ants. When they get food, they take it to the anthill, where it is processed and eaten by all members of the colony. Many species of ants exhibit a high degree of coordination and cooperation between “prey” (usually through pheromone transfer of information).
The "ant farm" mimics the evolution of complex behavior in artificial organisms. It considers organisms that live and reproduce in a relatively complex environment, with many senses (internal and external), and a large set of possible actions at any given time. Moreover, each ant has a memory, and, therefore, its behavior depends on its history. Throughout life, each ant is born, manages to make thousands of different decisions and actions, leaves offspring (this depends on the behavior of the organism throughout its life), and eventually dies.
By the way, the AntFarm model was developed in CM ++ based on the Connection Machine 2 supercomputer , which specializes in developments in the field of Artificial Intelligence.
UPD Article, in English. It more or less in detail describes the model of Collins and Jefferson - AntFarm.
- Part 1 (PolyWorld, Interaction between learning and Evolution)
- Part 2 (Grasshopper, AntFarm)
- Part 3 - coming soon ...