They use a utility function to drive their bot target selection. In one of their 2020 dev blog they mentioned that the utility function is a layer above the input values given by the enemy. A 2021 dev blog revealed that these input values are:
If the target is moving, and what the movement speed is.
Body position of the enemy. Prone is less likely to be noticed than crouched, and crouched - less likely than standing.
Distance to target.
If the enemy is within the field of view or not.
The sound of shots have very high priority in attracting AI attention, factored also by the distance to any flying bullets.
The output of this function give a “danger level” value of the enemy. Machine learning is used to train this utility function by curating the parameters of said function to improve the agent’s target selection. The bot will attack the enemy with the highest danger value or switch to attack them if they highest danger enemy is not their current target.
ig, but the issue is none of the old games really use ai ai, just things that simulate intelegence within set boundaries, what i have to reaserch sadly is true ai, which while super intresting, unfortunatly is super difficult to find with alot of the games ‘ai’ being mistermed, but i think that there is one aspect of the squad in this game which does incorperate actual ai which is very good
If you test it in editor. AI still has dynamic path finding. Which will dynamically changed path according to destroyable scenes. (Lots of older games have to “bake” AI pathfinding area)
But AFAIK, it’s still dev who is tweaking AI’s behavior, especially shooting and detecting sensitivity. Also such as making them afraid of water (which AI used to swim very often)
It’s still “human learning” players feedback and game’s AI log and change their behavior.
What “deep learning” I define is, AI and algorithm will change by itself after receiving input, and concluding human preference and change to meet the preference.