State of AI in games
AI in games is a divisive topic amongst many gamers. How good is it and how good can it be?
If we look at immersive real-time games, like the revered Deus Ex for example, there hasn't been all that much progress in the last 15 years or so.
Star Citizen is currently pushing the edge, and it needs to be seen what Cyberpunk 2077 reveals.
Mostly current AI in games looks at the actions that the player or an AI actor can perform, and assigns a value to each possible response. If the AI is more sophisticated, this evaluation is context-sensitive, so that different responses depending on the circumstances emerge. This allows you to build a fairly complex AI behaviour with a sufficiently large effort.
Whether this is good enough for your game depends. Some games require no more than basic response patterns. For example, if you shoot at an enemy squad member, he can take cover or shoot back. But if you look at something like the combat between two squads of 20 members each that Star Citizen plans, a game will hugely benefit from a much better organisation of all squad members. Somewhere there is a line where this AI approach is good enough and where it becomes insufficient. Often games design their encounters in a way that it requires frantic action, so that these types of deficiencies don't become obvious to the player.
But there is a huge potential for games that can take this to the next level. The limitation is currently how you arrive at possible responses. Only looking at the responses to an action and assigning these values, even context-sensitively, is intrinsically shortsighted. If you examine more closely the equivalent behaviour in the real world, typically you look at the actions that you can take and consider what your opponent can do. But you don't stop there. You see that your opponent has an idea of what you can do and this will impact his response. For example, if he would take cover behind an isolated obstacle, and all you would need to do to take him out is lobbing a grenade over this obstacle, it is a no go. Instead he could take a risk, trying to suppress you with a burst of fire and make a dash for a more safe position. And knowing this, you possibly would choose a different approach. It is a complex give and take. In AI terms it is a lookahead to determine possible responses to actions, and responses to these responses to construct sequences of actions and counteractions and arrive at a tree of variants to determine the best sequence and initial action. It is inherently recursive, because actions and counteractions depend on each other. It is also not easy to do because you don't know how far ahead you have to look, and the computation is expensive and grows exponentially with the lookahead depth.
This gives you an idea why these AI advances are difficult. But if you can solve it, your game will never be the same. What you would want, going back to the previous example, is that your sniper on a high vantage point knows that he doesn't need to shoot if the enemy advances towards a position that is difficult to defend against his comrades lying in wait hidden. That you have a highly organized squad that can adapt to threats smartly. Once you have achieved this, this intelligence propagates to other game elements easily. For example if you encounter the leader of another squad, before hostilities break out, and he knows that his squad can deal with your squad that is following behind you, he will not be intimated. But on the other hand, if your squad is better equipped and he can gauge this (via the AI), he might be intimidated.
What this requires is an AI that can arrive at sequences of actions and counteractions to choose a suitable behaviour, and using this AI logic to balance the values of actions. This goes for singular actions, like using a tool to open a backdoor to complex estimates like whether it is a good idea to engage an opposing faction.
This is the type of technology I am working on. I have a solution for the inherently recursive problem of finding the best sequences of actions and counteractions. And if you apply it to a gameworld as a whole, it structures all its elements coherently. You get an AI that can see all the interactions that can occur. From this point onward arriving at the value of actions is no longer a basic action response valuation. You get access to a whole range of new tools to craft realistic and nuanced AI behaviour. Naturally this approach can be applied to any subset or part of a game. Vice versa it is an excellent tool to integrate all parts of a game seamlessly.