The intelligence of non-player characters has long been a used as a selling point for new games by publishers likeEA. The player’s immersion is deepened when the characters that populate a games world react to them and each other realistically. Most modern games NPC’s all work in a similar sort of way, with the quality of the NPC being dependent on the level of complexity attributed to it by the developer.EAis now looking for a way to move this intelligence forward while cutting back on the complexity.
The player is the only one in a video game world with any true agency, asNPCs can only react to a player’s actions. These NPC reactions are decided by behavior trees, which are essentially flow graphs deciding what the NPC should do next. The perceived intelligence of an NPC is measured in how complex and varied these behavior trees are.
RELATED:EA COO Says Problematic Gaming Executives ‘Must Go’
The problem for developers is, for every possible player action a new branch needs to be added to the behavior tree for an NPC to seem truly realistic and reactionary. Not only that, but as more reactions are added, more computing power is used by every NPC.Cyberpunk 2077recently had to overhaul its NPCsto react more to player actions. But with a new patent filed by EA, NPCs could soon see a step forward in their intelligence.
The patent for is for “Readable and Editable” NPC Behavior. Essentially it describes a way of editing these NPC behavior trees on the fly, rather than having to increase the complexity of them to encompass all possible NPC actions. It also describes reading specific player data such as location and heath and using “reinforcement learning” to adapt that NPC’s behavior goal. But how close this system isto being used in future EA gamesremains unclear.
From the description this system could have many applications for NPC interactions, especially the reinforcement learning. NPCs could be able to not only recognize player actions in the moment, but remember actions players had done previously, outside scripted storyline choices. Or they could remember the last place they saw the player, or how much health they had, this would make for a very immersive game world. The patent also states that the generated behavior trees will be drawn up from this reinforced learning data. This means two different players interactions with every NPC in a game could be different based not only on their in-game choices, but play style.
EA’s proposed NPC AI learning system would be most interestingto see in the nextMass Effectgame, a series already full of diverging dialogue trees and different reactions based on player actions. Fleshing these NPC’s out and blurring the line between the good and bad reactions would be a step forward for the sequel. Outside of EA’s repertoire, a game likeHitmancould benefit from a learning enemy AI like this, where switching disguises won’t work on guards that have already seen the player.
