The endless saga of machines that surpass humans has a new chapter. An AI algorithm has once again defeated a human fighter pilot in virtual combat. The contest was the finale of the United States Army’s AlphaDogfight Challenge, an effort to “demonstrate the feasibility of developing effective and intelligent autonomous agents capable of defeating adversary aircraft in air combat.” ”
Artificial intelligence is capable of winning the human being
Last August, the Defense Advanced Research Projects Agency, or DARPA, selected eight teams ranging from large traditional defense contractors like Lockheed Martin to small groups like Heron Systems to compete in a series of tests in November and January. In the final on Thursday, Heron Systems emerged victorious against the other seven teams after two days of old-fashioned aerial combat, going one after the other using only nose guns. Heron then faced a human fighter pilot sitting in a simulator and wearing a virtual reality headset, and won five rounds to zero..
The other winner in Thursday’s event was the in-depth reinforcement learning, where the artificial intelligence algorithms they get to test a task in a virtual environment over and over again, sometimes very quickly, until they develop something like understanding. Deep reinforcement played a key role in the Heron System agent, as well as Lockheed Martin’s, the runner-up.
The algorithm has to adapt to reality in order to fly correctly
Matt Tarascio, vice president of artificial intelligence, and Lee Ritholtz, director and chief architect of artificial intelligence, Lockheed Martin, told Defense One that trying to get an algorithm to work well in air combat is very different from simply teaching software to “fly”, or to maintain a certain direction, altitude and speed. The software will start with a complete lack of understanding of even the most basic flight tasks, Ritholtz explained, putting it at a disadvantage against any human, at first. “You don’t have to teach a human that he shouldn’t crash to the ground … They have basic instincts that the algorithm does not have“In terms of training.” That means dying a lot. Hitting the ground, a lot, “Ritholtz said.
Tarascio compared it to “put a baby in a booth“.
Overcoming that ignorance requires teaching the algorithm that there is a cost for every error but that those costs are not equal. Reinforcement comes into play when the algorithm, based on simulation after simulation, assigns heavy costs to each maneuver, and then reallocates those loads as experiences update.
When you provide general rules, you limit their performance. They need to learn by trial and error, “Ritholtz said. Ultimately, how quickly an AI can learn, within a defined area of effort, is not disputed because it can repeat the lesson over and over again, on multiple machines.
A curiosity: this air combat system, its developers explained, does not need a lot of power, and is capable of working on a chip similar to NVIDIA Tegra.
This is not the first time an AI has fought a human fighter pilot in a contest. A 2016 demo showed that an AI agent nicknamed Alpha could beat an experienced human combat flight instructor. But Thursday’s DARPA simulation was arguably more significant, as it launched a variety of AI agents against each other and then against a human in a highly structured framework.