Golem Mancer: Layered A.I

For one of my game projects; Golem Mancer, I've been toying with a type of enemy A.I structures that I haven't really encountered before. Video game A.I is substantially different from the now ubiquitous idea of A.I like Large Language Models that power ChatGPT, Grok, etc. For video games. A.I is more about the illusion of intelligence rather than the emulation or simulation of intelligence. In this case, I'm looking to modulate the difficulty of my Golem Mancer using layers of artificial intelligence, rather than using more standard tools.

In many games, difficulty is fundamentally an issue of maths. Many Bethesda games are notorious for using overly simplified difficulty implementation. Changes between one difficulty and another can be as simple as increasing the Hit Points of an enemy and halving the attack power of the player. And while this is fully functional, and in many cases works out well, it is far from a robust system.

Certain types of games have design philosophies that can play a big part in how difficulty is treated. If I look at games like Frostpunk, Oxygen Not Included, or Factorio. For example have aspects that can be toggled on or off, randomization, and sliders that dramatically affect not only the difficulty but the balance of gameplay. Since these kinds of options give players more agency they are the gold standard for most of the games that support these kinds of mechanisms.

Unfortunately neither of these provides a template of even a foundation for using A.I as the crux of game difficulty. For that I took cues from Real-Time Strategy (RTS) games like Westwood Studios’ Command & Conquer: Red Alert, Age of Empires, and Civilization. These kinds of games, while not fundamentally working differently at higher difficulties allows enemy A.I to play closer to their full potential, closer to mathematical perfection and ideal strategies, with different priorities and faster reaction times. This makes it closer to what I’m looking to implement with parameters on the A.I controllers being the notable point of interest.

That being said there are lots of other game A.I’s not to be overlooked. I’ve spoken about the more advanced A.I in Alien Isolation a number of times which implements two sets of A.I stacked for improved simulacra of intelligence. There are also other sophisticated A.I’s in genres like:

• 
  

  Racing (Gran Turismo, Forza Motorsport Series)

  
  

  
  Beyond classic “rubber-banding” on vehicle performance, modern racing games employ various means of making enemies follow more optimized routes, and superior handling, and more tools to adapt to the player(s) driving style.
  

  
  



• 
  

  1st & 3rd Person Shooters (Halo, F.E.A.R, Left 4 Dead)

  
  

  
  Enemies will use cover and tactics more liberally. In some cases enemies will work collectively even sharing data in realistic ways like calling out to allies.
  

  
  




• 
  

  Strategy (XCOM, Total War…)

  
  

  
  Tactics become far more pronounced and the kind of induced vulnerabilities are set aside. The amount of abilities and data are increased to give the enemy teams more dynamic reactions and broader strategies.
  

  
  

I still found myself coming up short in terms of decent models to work off of. Everything I found was to far off the mark or not really close to what I needed.

I wanted the enemy AI to work collectively across different spectra based on the difficulty setting and for that to manifest as different levels of persistence. By persistence I mean that I want the A.I layers to last longer and longer, the easiest difficulty will support an A.I that only last while the player is on the same stage (each level is made up of several stages) and “learns” only during the course of that single stage.

As I tentatively designed my AI, my plan is to apply a separate set of A.I depending on the player’s selected difficulty to each layer. These A.I’s will share information vertically, with information and something akin to experience points percolating up from the lowest level A.I and the higher level A.I’s reciprocating with higher level skills that effect enemy tactics.

This should; fingers crossed, should if the implementations goes according to plan make the combat experience more dynamic overall. Enemies will on the one hand seem like that are actively adapting to the players strategies with counter-tactics as well gradually feeling like the enemies are predicting players rather than just reacting.

This will be accomplished using basic data crunching on the player. Looking at just the easy A.I, it will be gain experience as a function of time

and unlocks node that determine what it can extract in terms of useful data from across the battle field i.e from the players relevant data (golem stats, golem skills, golem abilities, etc.). This data is used to alter the enemy team’s tactics such as what target they will prioritize, what variety of units will be spawned, and when to use abilities to counter the player.

In tern the Medium and Hard A.I’s will accumulate experience in the same way but since they are persistent across an entire level and the entire game respectively they hold onto their experience longer. As such the skill nodes they unlock are more potent and persist.

They also take a different form. Because information about the player’s side is also sent to the A.I’s with increasing levels of detail the higher level A.I’s will process that information and form more advanced strategies using this information. For example if the player favors a Golems with a certain profile the enemy; once using more sophisticated tactics, will stack the deck in their favor. A player may find themselves walking into a stage where most if not all spawned enemies are strong against their lineup of golems.

The EXP style leveling up of enemies will be a serious balancing issue not only to make the game difficult within the margins of challenging but not punishing but also not too advanced that enemies don’t feel to artificial, just…smarter, as if a skilled tactician is across the table studying the player.

Finally, the third layer of the A.I; the hardest, will persist for the entire game, and while it will be hard capped at a certain point. This will essentially turn the game into a PvP with an active enemy A.I working against players the entire game. The desired effect of this, of course, is to make sure that the player must continue to adapt as the enemy gets progressively better at countering any given strategy. Hopefully this will have a couple of nice knock on effects, limiting min-maxing, as well as emphasizing all the strategic options and open up game.