Nothing bothers us, as
coming after a terrible storm
events dead the rest idle —
the clarity, where there is neither fear nor hope.
Mary Shelley, “Frankenstein”
This year the Pentagon is embarking on the most ambitious program of “training” of military vehicles — on the final stage of the electronic Neurocomputers unmanned ground vehicles and aircraft will be able to independently make decisions in a combat situation.
The American authorities have already approved the corresponding budget. The project, as one would expect, will be engaged in Management of perspective research projects of the Ministry of defense or DARPA. It is expected that the research programme in the field of military unmanned systems the U.S. military will spend about $ 10 billion.
DARPA has already successfully implemented the development of a single combat information management system, which will be the same for all types of military drones. Such a system is a complex analytical neurocomputer, which summarizes all obtained from the field data and allows a user to exchange information between the unmanned devices to coordinate their work in the context of military operations.
As a result of the implementation of the programs DARPA, drones will be able to decide on the strike depending on the target types without operator intervention. Such drones learn to act depending on the priority objectives, to distinguish friend from foe, act completely Autonomous. Drones will be able to set priorities and to act in a specific attack or a specific spontaneous battle scene. Neurocomputers will learn to operate it like the real aircraft pilots or tank crews. The importance of this project can not be underestimated.
In this article we consider the features of the neural network intelligent devices control onboard systems of unmanned aircraft. Unfortunately, in the Internet there is practically no publications on this topic. Today we will try to fill this gap. For a better understanding of the material read the material presented in the previous article.
Modern unmanned systems.
The concept of “Autonomous drone” today is very vague. Most models produced today have a certain level of autonomy, can move along pre-programmed routes, followed by intermediate points, automatically use emergency modes, for example, when connection is lost or battery is low. The most advanced capabilities of these devices today — detection and collision avoidance, formation flying, and the execution order of the tasks.
However, in the world today is developing effective electronic warfare systems, jamming. Rapidly increasing capabilities of the hardware which allows the use of drones far beyond the possible areas of their reliable control. Thus, in the DARPA report States that autonomy is an increasingly important factor in the development of unmanned technology.
DARPA also recommends to combine the efforts of designers and specialists on artificial intelligence (AI). Moreover, DARPA assigns a special role to introduce new technologies in production, and the creation of new fixation systems for the new weapon. The idea was called “the factory in one carton”. That is, all consumables and the necessary tools, including CNC machines, processors, on-Board electronics, cables, etc., shall be transported directly to the place of combat operations and be stored in multiple modified cargo containers. Also in the shortest possible time should be prepared by the appropriate staff.
The neural network device (the electronic brain)
We will now consider the hypothetical configuration of artificial intelligence fully Autonomous combat system. Of course, yet such machines are not created, and we can only imagine based on logic and knowing the basic functionality of these devices, what will be the architecture of neural computers.
The latest devices are designed to solve many more combat tasks than the existing existing examples. The decision of many military missions require installation on drones fundamentally new systems to obtain and record a large amount of information, including audio and video format. Thus, the load on the artificial intelligence of the apparatus will be sufficiently serious only because of the volumes of incoming and recorded data. As we said in the previous article, the function of pattern recognition is best executed of the so-called convolutional neural network.
At the same time, to recognize a different type of information be more appropriate recurrent network. There is also Bystrovskaya single-layer neural networks, which are more suitable for efficient operation in combat. Accordingly, the neurocomputer of the drone must exist a certain set of configurations of neural networks.
Hypothetically this can be achieved in two ways. First, it is possible to combine several different types of Neurocomputers in a single system. This method is the most obvious. However, really promising is another way to solve the problem. We need to create such an artificial intelligence, which alone will make the decision about the reconfiguration of the neural network depending on a solved problem, whether it be observation of the terrain, operation
air or ground combat, condition monitoring of onboard systems and so on. This is a very economical solution, requiring, moreover, communication between the various modules, which, of course, complicate the whole scheme in General.
Now consider a hypothetical operating mode of the Autonomous neural apparatus during the execution of combat missions. In the onboard device control digital signals from the sensors object of control goes to the digital signals. At the same time, analog signals from the sensors object of control goes to the analog signals, where they are filtered and analogoue-to-digital conversion.
Then the processed signals are transmitted in parallel to the inputs of the neural network (BNS — unit neural network) where they are processed, after which the output comes to on-Board recording device (BUR) and the automatic reconfiguration of onboard equipment (houart BO). Varos BO performs certain actions in an emergency actions to disable the faulty unit, switch to use a backup communication channel for the safe completion of the flight or the issue signal to alarm system device with the transmission failure message and coordinates to the ground control station.
Part of the ground system is designed only for manual configuration of the on-Board neurocomputer for given parameters (minimum and maximum values of input and output signals, parameters of the architecture of the neural network), the onboard control part after it is setup and planned maintenance in the terrestrial environment. All solutions neurocomputer takes independently.
The core of the airborne control of unmanned combat vehicle is, of course, the neural network block (BNS). The converted signals are processed in the neural network module, which operates according to the following algorithm:
to the input of the neural network is fed a current signal;
— calculates output signals of the first neural layer;
— calculated layer by layer, output signals of the hidden and output layers.
The signals of the output layer are the resulting signals of the neural network. In the module output signals are resulting signals of the neural network is transformed to the sj , convenient for users.
In the memory module stores information about the parameters (minimum, maximum) input and output signals, the architecture of the neural network (weights, number of layers, number of neurons), obtained in the result of learning BNS with the ground part of the control system.
Of course, an important parameter side of the neurocomputer is the ability to self-education.
The process of learning BNS involves a series of steps:
1. The definition of neural network architecture:
— the number of inputs and outputs;
— the number of layers and neurons in each layer;
— the values of displacement signals to neurons.
2. The channel setting module preprocessing the signals:
— enter the minimum and maximum possible values of input signals;
— enter the minimum and maximum possible values of the given signals.
3. The channel setting module output signals to convert the signals of the neural network in the form required.
4. Training the neural network.
The last stage of learning BNS is the most time consuming from a computational point of view. Currently developed various methods for training neural networks that allow to automate this process. The most widely known method of error back propagation. On the basis of it is possible to build a very efficient learning algorithm of neural network that represents the following:
a) select the value of the learning step 0< p < 1;
b) sets the value of the maximum acceptable error E max for any neuron of the output layer of the neural network;
in) initially selects the weights of connections in the neural network and the initial values of the weighting factors of displacement. Their values can be set randomly;
g) on the basis of known expertise, knowledge formalized in the form of technical
description and operating instructions of the control object, the data obtained in the test object are determined by training sets of the K examples of the converted values of the input and corresponding output signals of the neural network;
d) for a given value of the input of the neural network first example of the training set to calculate values on the output of each neuron in the first layer and all subsequent layers alternately.
e) defines the error values for each neuron in the output layer and all previous layers alternately in descending order of number of layers.
f) an adjustment of the weighting coefficients of neural connections between the first layer and the inputs. The same operation is performed for offsets of neurons in the first layer, alternately between all subsequent layers, and for offsets of neurons in subsequent layers.
h) then calculations are performed for each of the remaining examples of the training set;
s) after the current cycle of training on all examples, the calculated error value at the output of each neuron of the output layer of the network for the entire training set. In connection with the importance, from the point of view of security of the device, neural network device monitoring is necessary to ensure the accuracy of the signal is not below the value set at each output of the neural network, so it is advisable to use the stopping rule of the learning process of the neural network based on the comparison of calculated E (K) with a given maximum error E (max).
— if E(K) > E max , the training is repeated;
— if E(K) £ E max , then the neural network is considered trained and ready for intended use.
Naturally the question arises about the possibility of practical implementation of such a neural network device. There are two options for the implementation of the neural network software (programs on conventional computers, which are called nanoemulsion) and hardware. Although software implementations outweigh the hardware, however, the share of hardware implementations of neural network devices in the total number of global technological developments is growing steadily. The greatest development in this direction is observed in Japan, where there is a continuous development of neural computers for military applications.
Prospects of realization.
The first is to emphasize that on the battlefield as a result of successful implementation of this technology will no longer be killed the crews of aircraft and ground combat vehicles. There will be underwater unmanned vehicles, warships will receive the “electronic brain”. Even those systems that will continue to be managed by the person, will receive, as a spotter, Neurocomputers.
Invaluable is the introduction of robotic systems for the bomb squad. A standalone device can cope with the task faster, the neurocomputer no emotions, the probability of error is minimized. The Neurocomputers can work independently where the operator of the robot simply cannot communicate with the device.
Happens miniaturization of military vehicles, as such do not require sophisticated life support systems, emergency ejection and much easier living pilot. Accordingly, there will be an increase in the number of ammunition.
Neural networks can be applied in devices with a nuclear engine, since the radiation has on such such a devastating effect on the human body. Accordingly, it will be able to create long-term patrolling the airspace of the missile-carrying bomber, which to no living crew, and all decisions will be made by the neural computer.
As we can see, technology in the military industry are developing very rapidly. In 80-ies of such a system was considered a prospect for the distant future. But the future came sooner than we expected. And as always, the future comes to us on the platform of military developments.
Tanai Cholhanov, especially for News Front