This month I thought I’d change things up a bit and talk about one of the backend systems in Revenge of the Firstborn, specifically the AI system. RotFb uses an AI approach called utility AI. In utility AI, each action that an agent can take is given a “utility” score where utility refers to how happy the actor would be if it took the given action.
To determine the action with the highest utility score, the game loops through a series of a couple dozen different potential actions, ranging from ending the NPC’s turn, to casting a spell or making an attack. Each action has one or more decision inputs, each of which has a numeric value for when the condition is true and numeric value for when it’s false. Those values are then added up to get the final utility score for an action.
Let’s take an example of drinking a healing potion. To get the healing potion utility score, the game has several inputs that can raise or lower the final score. They are:
- Does the NPC have a healing potion in their inventory? Naturally, if they do not, the utility for this action needs to be set to a very low score. Since I know the score for ending the NPC’s turn is 0, I give this input a value of -100 to ensure that no matter what other decision inputs modify the action’s overall score, it will still be below 0. If the NPC does have a healing potion, the score is unchanged because simply having one in their inventory has no bearing on whether they want to use it.
- The second-most important input involves evaluating the NPC’s health. If they are at full health, this check adds nothing to the score, however if they are below, say, 50% we increase the score so the final score for this utility is higher than the baseline of 0. Let’s say this action increase the utility score to 25.
- For this example, we’ll include one last decision input. Is the agent close enough to an enemy that the enemy can make an attack of opportunity on them if they drink a potion? If so, we reduce the score by 10. This would make the action’s final score 15, meaning it is less likely to be chosen, but not impossible to be chosen.
Let’s say that our hypothetical agent has a potion, is low on health and is not in danger. This would make our Use Potion utility have a final score of 25.
We don’t have any enemies close by, so the attack utility is low, perhaps 10.
However, the ai actor has a fear effect. The fear effect has a very high utility score because fear supersedes any other actions the agent could take. The Run Away in Fear action has a score of 75.
So, we are left with the following utility scores:
- Use Healing Potion – 25
- Attack Enemy – 10
- Run Away in Fear – 75
Making the clear winner Run Away in Fear. The actor will attempt to find a place that is far away from the source of its fear and run to that location.
The game has several baseline utility action collections – it has one for average intelligence agents, non-intelligent agents (undead) and even a few specific ones such as dragons. This helps give agents different behaviors as appropriate. For instance, average intelligence creatures are smart enough to attempt to flank you in combat, but non-intelligent ones are not. Creatures can also add new actions to the baseline collection. For example, the medusa has an addition action for using its petrifying gaze attack.
The game uses a similar scoring system for what type of an attack an agent should make. For instance, a trip attack gets a higher score if the agent is bigger than its target and it has the Improved Trip feat. Grappling is more likely if the agent is clearly stronger (a very large difference in strength score) and if the target is a spellcaster who would be largely neutralized by being unable to freely move their limbs.
The AI also scores spells in this manner. Each spell the agent knows gets a score based on how many targets it can hit, whether or not there are allies in the area of effect (assuming the creatures cares about its allies) and so on. In order to make the choice of spells a little less predictable, each spell with a utility score within 10% of the highest has an equal chance to be chosen. This gets us one of a few viable spells but also excludes all spells that are clearly not applicable to the current situation.
Hopefully you’ve found this little peek under the hood of the engine interesting. Keep an eye out for more details in future updates!