From chess to poker to deepfake videos , AI is beating humans. Now, as if that were not enough, another defeat arrives: the artificial intelligence has beaten an F-16 pilot in a fight between fighters .
We are obviously talking about a virtual competition, held in the terrain of the AlphaDogfight tournament organized by DARPA . The algorithm beat the professional driver 5 to 0.
When AI beats an F-16 pilot
Who knows what it feels like to be beaten by artificial intelligence in the field where you have years of experience. This is what happened to a US Army pilot who was defeated by the AI in a fighter race .
The AlphaDogfight is a tournament organized by the DARPA (Defense Advanced Research Projects Agency) with the aim of identifying the best artificial intelligence system in the military field. After several matches between AI, the algorithm of the software house Heron System prevailed and found himself facing a real driver in the final round.
The pilot, who has recently become an instructor and with more than 2000 hours of combat on the field, could not help but accept the crushing defeat of 5 to 0. The two contenders challenged each other in five different scenarios with basic combat maneuvers. Each time the AI was able to execute the correct maneuvers and shoot down the pilot .
The AlphaDogfight is part of the DARPA ACE (Air Combat Evolution) program , which aims to improve autonomous combat in the air, relieving pilots from as many tasks as possible .
"In an air-to-air collision a single human pilot could increase lethality by managing multiple autonomous platforms at the same time," explained a representative of the American agency. "This would change the role of the human being from a simple operator to a mission commander".
" The human focuses on what they do best , such as strategic thinking," said Timothy Grayson in the face of humans' fears of sunset as they fly planes, " and the AI handles the rest as if it were a weapon. evolved ". A collaboration and not an exclusion, therefore.
Behind the winning AI is a specific learning approach called reinforcement learning or reinforcement learning . Among the main paradigms of machine learning together with supervised and unsupervised ones, this technique is used for decision-making-sequential problems.
This particular training process is based on the concept of reward : the algorithm studies the environment where it is located, in which every action made by an agent changes its state and causes feedback . If the feedback is positive, a reward is awarded (positive real value), while if it is negative, a penalty is assigned (negative real value). A reinforcement function is used to measure the degree of success of an action against a goal.
The function works on the external environment by representing it as a vector of characteristics . An environment, for example, can have three binary characteristics: x1 = rain, x2 = clouds, x3 = wind. The combination of these values gives rise to the vector with the different states in which the environment can be found. The purpose of the agent is to maximize the reinforcement function .
Since the decisions of each step depend on those taken previously, the system needs to "remember". Memory is represented by the Knowledge base , in which a certain numerical value is associated with each state Xi. The agent, therefore, will repeat over time only the most profitable actions.
The data collected is used to generalize the decision-making model in order to make choices even for states where the system has never been found. The choices are taken for "similarity" with respect to a profitable state already experienced .
The article Artificial intelligence beats an F-16 pilot in AlphaDogfight comes from TechCuE .