Developer Keiwan Donyagard (Keiwan Donyagard) introduced Evolution – a small simulation game that allows you to monitor the work and development of artificial intelligence. Using a set of bones, muscles and joints, you can create small creatures and give them the task of learning how to jump, run or walk, and then follow their evolution. Simulation can be tested on the site.
The task of any kind of machine learning is to learn to perform a task in the course of continuous attempts. The type of training depends, for example, on the architecture of the chosen neural network: for example, in generative and adversarial systems, the network generator generates objects, and the estimator network compares them with the “gold standard”, and then sends the generator to redo the object until the best result is achieved. Usually, when analyzing the work of machine learning, one can look at the weights-the estimates that are given to the solution obtained. The evolution of artificial intelligence in the learning process can now be considered in every detail – with the help of a new simulation game.
Each created creature is equipped with three types of details: bones, muscles and joints. Joints are necessary in order to connect the bones together, and the muscles regulate the movements of the bones due to contractions and stretches. In the process of “evolution”, a brain is added to the existent, a neural network that is trained on the movements of the being. The only part that the brain can control is the muscles, which should be enough to move all the bones (that is, the bone can not be connected to the other bone only with the tendon).
In the process of “evolution” the creature has several copies that move the muscles in a random order. With continuous stretching and contraction of muscles, the system learns by analyzing a certain set of parameters: the distance of the creature to the ground and the number of points of contact, its direction of movement and speed, and also its location in space. Then, depending on the task, the system selects the two most successful candidates for further propagation: the corresponding parameters are then used to create new creatures. The process is repeated until the task assigned to the creature is performed as correctly as possible.
With the help of such a simulation, as Donjagard notes, one can follow how the neural networks are trained: first randomly performing actions, and then adjusting their behavior based on the analysis of the parameters given to it. In addition, simulation is also an excellent way to learn more about the structure and functioning of devices that work with artificial muscles. In addition to the web version, there are also desktop versions for different operating platforms.
To effectively teach artificial intelligence, think of other ways. For example, last summer, American researchers were able to speed up the learning process of the robot, forcing another robot to take away items from it.