View source for Machine Learning Techniques for Autonomous Mutual Separation in Fighter Formations

{{Masterproject
|Master name=Human Ambience
|Student name=Marleine van Kampen
|Project start date=2014/02/03
|Project end date=2014/08/31
|Supervisor=Mark Hoogendoorn
|Second supervisor=Jan Joris Roessingh (NLR)
|Company=NLR Amsterdam
|Thesis title=Machine Learning Techniques for Autonomous Mutual Separation in Fighter Formations
|Finished=Yes
|Thesis=Thesis.pdf
|Poster=Posternaam.pdf
}}
<b>KIM 1: </b>

The NLR (National Aerospace Laboratory) in Amsterdam uses a simple F-16 desktop simulation to research machine learning techniques for application in simulations of virtual fighters (intelligent agents) engaged in air combat. The simulation covers situations of one-versus-one (1v1) combats and two-versus-one (2v1) combats. Simulations that can handle scenarios with larger numbers of fighters in combat are under development, like two-versus-two and four-versus-four. 

Generally, aircraft in a formation need to co-ordinate their flight paths in order not to collide with each other. The virtual fighter aircraft in the simulation should also exhibit this co-ordinated behaviour. The current simulation included only a simple flight model. To implement the machine learning techniques and see the realistic behavior of the separation algorithms a more realistic way of aircraft movement is required. 

Tactics are based on mutual separation between flight members in the formation as well. One of the general tactics to have a higher chance to win the air combat is to fly in formation. By having knowledge about their relative location, flight members can fly in formations. By using machine learning techniques formations can be learned to the teams and the most efficient one can be selected for a situation. The learning techniques may even be able to form more efficient formations then are currently in use. 

The research mainly focuses on the question: “Which machine learning techniques are applicable for autonomous mutual separation?”. Furthermore, the following question will be answered in the research:

•	What are the advantages and drawbacks for each of the machine learning techniques?

•	What kind of flight model can be implemented in the simulation to enable realistic aircraft maneuvering in case of collision avoidance?

•	Can the chosen technique be used in other contexts, such as drones and RPAs?

•	Can the chosen technique help to form formations in the simulation? And can it generate new types of formations?

<b>KIM 2: </b>
 
Training pilots in an simulation environment as realistic as possible is a challenge for the research field of Artificial Intelligence. Within simulation environments, as the NLR (Dutch National Aerospace Laboratory in Amsterdam) develops, fighter pilots are trained how to handle air combat situations. These environments should reflect the real world and therefore should use the same knowledge to operate tactical within an air combat. The knowledge can be added within scripts and scenarios but the tactics are too many to use only expert knowledge. The simulation should be able to learn from the past in what choices are best to be made. Previous research showed already some of the learning behavior for situational awareness and selection of rules, but one not to forget concept is missing. The separation between the two teams to avoid collision, and the separation within a team to fly in formation is a requirement for a realistic air combat simulation. These concepts should be able to be learned with Machine Learning algorithm. Reinforcement learning uses experience from the past to adjust the behavior. Multiple algorithms that could be applied to the air combat simulation, namely Neuroevolution, BDI evolution and Dynamic Scripting. The current research highlights the use a Neural Network in combination with evolution of the weights of the network, known as Neuroevolution. An set of experiments is conducted to research whether Neuroevolution is applicable for the domain. The experiments for conflict detection and resolution show that the use of a Neural Network for calculation is applicable in combination with the learning of the weights by evolution. As the experiment for Autonomous Mutual Separation has the same modelling, it is applicable as well, although some more efficient coordination within the team should be researched.