To pass the evolutionary path from amphibians to mammals, animals took many millions of years. In the course of evolution developed and their brain. The existing neural structures were gradually added to a small group of neurons. As a result, the complexity of the brain is gradually increased, “keep up” with the development of limbs and sensory organs.
A “brain” (computer software) Conventional robots are not able to develop. If, for example, a robot limbs add new sensor type, you will need to complete the processing of its software, which is expensive and time-consuming. To make a humanoid robot with human-like behavior, you want to use more and more sensors. It is imperative that the software of the robot itself has grown in complexity – in the same way as it does the brain biological beings.
Engineer AI Christopher McLeod and his colleagues at the University of Robert Gordon (Robert Gordon University) to Aberdeen (Aberdeen), United Kingdom, have created a robot that mimics biological evolution.
The robot is controlled by the neural network: a computer program that mimics the learning process of the brain. The neural network comprises a plurality of interconnected processing nodes, which can be “trained” to perform desired actions.
For example, if you need to maintain a balance, and the robot receives input signals from the sensors, he topples over, the robot will move his limbs in an attempt to keep his balance. Such actions provided limbs controlled importance ( “Weighing”) of the input signal for each node.
Certain combinations of input signals from the force sensor assembly to generate a signal, such as turning on the motor. If the robot stood on their feet, the combination of signals is maintained.
If the robot has fallen, it will make adjustments in their actions and next time try to do something different.
McLeod Group went further and developed an incremental evolutionary algorithm (IEA), capable over time to add a new part in the “brain” of the robot. They started with a simple robot the size of a book with two rotating prosthetic legs, which could turn the engine 180 degrees.
Then they gave the shestineyronnoy robot control system, the original team – for 1000 seconds to move and not to fall as long as possible. To accomplish this task, we installed a computer program, which is to develop the best way to travel. The robot always first to fall, but then he began to move forward and not fall while driving in a straight line. In the end, the robot learned to move like a tadpole.
When the IEA recognizes that its evolution is no longer increases the robot’s speed, it captures the state of the neural network, depriving it of the opportunity to develop further. This network already knows how to operate the robot-prosthetic legs – and it will continue to do so and on.
If now in the leg prosthetic robot adds “knee” any other, even-paced, the robot would not be able to manage such a modernized foot.
But unlike conventional evolutionary algorithms, an incremental evolutionary algorithm feels his sudden inability to fulfill its primary command. Therefore, when at his feet-prostheses appear “knees”, the program “knows” that she would have to learn to walk again.
To do this, the program automatically connects to their “fresh” neurons to learn how to manage his new legs. In the end, we are developing a robot motion, motion resembling salamander legs.
After starting the command is successful, an incremental evolutionary algorithm captures the state of the second neural network.
If the rear of the robot’s legs to add the program again adds more neurons, and this time the robot develops motion gallop. When a robot equipped with a camera, he has learned to go to the light or to avoid light