To put the Alpha Dogfight Trial in perspective, I found this[1] video by a former fighter pilot quite informative.
For me the key point was that the rules were such that simply holding a "cone of fire" would "damage" the opposing fighter, no actual bullets were simulated.
This meant that the optimal way to "win", and thus what the AI optimized towards, was by taking approaches that in real life with actual bullets etc have a low chance of success.
Coupled with perfect situational awareness, AI got fed exact position and momentum for each plane among other things, this really tipped the scale towards the AI.
Heron System comments that they're moving towards a more realistic scenario next with simulated sensors and weapon systems. Will be interesting to see how well it fares.
1 comment
[ 4.3 ms ] story [ 8.9 ms ] threadFor me the key point was that the rules were such that simply holding a "cone of fire" would "damage" the opposing fighter, no actual bullets were simulated.
This meant that the optimal way to "win", and thus what the AI optimized towards, was by taking approaches that in real life with actual bullets etc have a low chance of success.
Coupled with perfect situational awareness, AI got fed exact position and momentum for each plane among other things, this really tipped the scale towards the AI.
Heron System comments that they're moving towards a more realistic scenario next with simulated sensors and weapon systems. Will be interesting to see how well it fares.
[1]: https://www.youtube.com/watch?v=ziCQqmEllZo