Modern Naval Warfare | ANN for WMZ prediction
Alpha Defense Exclusive
– Admin Alpha Mariner
With the passing of every hour, every day, every week, the necessity for incorporating autonomy into modern naval warfare is rising. Such necessities are first driven from prowess acquired in technological development and later on due to the emanating threat that nations perceive given their adversaries’ progress. In the first two articles of this series of Modern Naval Warfare we touched upon the concept of USVs and how the rise of ANN, machine learning and artificial intelligence has facilitated the transition of role for ships to motherships. There we remarked upon quite a few existing and in-development USVs.
Following that, we analysed the CVX aircraft carrier design for South Korea by Daewoo Shipbuilding and Marine Engineering (DSME). We remarked on how the design and general arrangement changes were made with focus on modern naval warfare. And finally in our previous article we remarked on US Navy’s vision to make use of USVs as a complement to Distributed Marine Operations and Distributed Lethality (DMODL). There we have touched upon the two USVs that the US Navy is working on to bring it into full fledged operations,how they are trying to overcome the challenges faced especially the processing of humungous volume of data and its translation to operational decisions, despite the availability of best algorithms.
From here on, articles on modern naval warfare will start getting technical. So we advise the readers to go through the hyperlinks that we provide in between the articles. Ofcourse, we shall be staying true to our motto of ‘Simplified Defense’ and break down concepts to the layman level. However writing down every basic thing would lead to diversion from the main topic and so the article shall be so as to provide a big picture. Specifics can be found in the hyperlinks.
ANN Fundamental Big Picture
The concept of data science is to collect huge amounts of raw data and process them before an artifical intelligence can use it. Think along the parallels of a human brain. Not all of the raw data can be used most of the time. This is because the data so obtained is usually the result of an experiment or statistics done that in most cases contain errors due to various factors such as measurement flaw, reporting etc. Another aspect is the data need not be numeric always. It can also be texts. Even in numeric data, the range of each variable differs and hence analyzing the data in same scale becomes a problem. Huge volumes of data contained hundreds to thousands of variables cannot be represented in 2D or 3D graphs alone. Higher dimensions cannot be perceived by humans visually. Hence it becomes necessary to transform the data into 2D or 3D before visualzing it. This is how data science helps. The fundamentals of data science algorithms lie in probability concept and logic. Hence we advise the reader to have some basic idea on probability or go through this link.
Now once the data has been processed suitable enough to be trained on, a machine learning models (nothing but a piece of code; the logic is called algorithm) will be fed with the data and the model computes certain parameters of its own based on the data. The model learns by itself on how the given data behaves. These parameter values are now used to predict / estimate / classify the existing problem at hand with a particular level of accuracy. The accuracy levels depend on how much data, how well the data is preprocessed and the tweaking of certain parameters within the model. Read more about machine learning here.
Now once a solution to the problem is found, it is used by another set of code that translates the solution into an operational decision. This encompassing code is the artificial intelligence (AI). Note that AI at it’s heart comprises of a machine learning (ML) model. An AI figures out the solution without us having to tell what is to be done. Read here about artificial intelligence.
There are many machine learning models out there, information about which can be found here. However in this article we shall be focussing on a neural network based model. Neural networks are machine learning models based on how neurons in the human brain work. Feel free to check out this link to know about neural networks and it is advised that the user further read upon ANN i.e Artificial Neural Network here.
ANN for WEZ computation
WEZ is the weapon engagement zone for a missile. Conventional algorithms used for simulated include
- One where there are background flyouts of the missile model continuously during the simulation
- Second where simulated flyouts are executed offline and later on creates a lookup table using the already generated truth data. In such cases interpolations are done between simulations on truth data in order to generate data not available in table. Usually multiple adaptive regression spines or projection pursuit regression models are used.
A much later development was use of ANNs for ascertaining the WEZ of a missile. This is done by using the practice missile firing data as the dataset. This data is then used to train a multi-layer ANN with the Bayesian algorithm. These algorithms are embedded within the Fire Control Computer (FCC) onboard the ship. The ANN model tries to generalize patterns based on the data with which it is trained. Training an ANN is computationally intensive but once done, the said model can be used n number of times. ANNs also predict faster thanks to the use of Stochastic Gradient Descent Algorithm employed. A feed forward multilayer perceptron (MLP) is the type of ANN used wherein learning happens by the adjustment of connection weights (the parameters we were talking in above paragraphs) based on the error that is backpropogated through the layers. The objective of the ANN is to adjust the weights until the error is minimized.
Typicaly each input to the MLP is an engagement scenario detail comprising of WEZ for each scenario.The MLP then trains on it and later when a test scenario is provided, it will provide real time WEZ for it. Some of the inputs typically include
- the maximum launch range based on the current engagement when the missile will hit the target for a minimum Mach number
- the No escape Zone limits for AAMs
- the target coordinates
- target speed, heading, altitude
- positions relative to the ship say the pitch and angle off sight
When it comes to the activation function used in the final layer, those are either threshold functions or sigmoid functions such as logistic function or the hyperbolic tangent function. When it comes to temporal i.e time varying data, recurrent neural networks i.e RNN are used. It is important to understand that there always is some error associated after training. This is more or less optimized so as to not to overfit or underfit the model, thereby helping the model retain as much generality it can. The error is typically expressed as average error over training as well as the average error during test. It is important to understand that the data set must be large enough to minimiz the error and to facilitate the model achieve as much generality it can.
Deep Learning NNs
A very recent development has been the use of deep learning based neural networks for missile guidance. One of those guidance laws that replace the existing proportional navigational guidance law is the deep neural network based guidance law. Much like the ANN, this also takes a supervised learning approach. The outputs considered for performance evaluation include the hitting rate and the energy function. The upcoming article in the modern naval warfare series will touch upon these models.
Copyright policy : This article is exclusively written for Alpha Defense ( https://alphadefense.in/ ) by Admin Alpha Mariner, the author. The content may be referenced but this article or its sentences either in full or in part cannot be exactly copied or republished in any form be it article, audios, videos, infographics or clippings without the written consent of Admin Alpha Mariner, the writer. The content if referenced, due credit must be given to Alpha Defense i.e https://alphadefense.in/. Violation of our copyright policy will be subject to legal action within the framework of Copyright Act, 1957 India.