Smart NPCs have long been used by publishers such as EAs as a selling point for new games. When characters in the game world react realistically to them and each other, the player's immersion deepens. NPCs in most modern games work in a similar way, i.e. the quality of NPCs depends on the complexity given to them by the developers. EA has now filed a new patent to find ways to advance this intelligence while reducing complexity.
For the video game world, NPCs can only react to the player's actions, which are determined by behavior trees, and the perceived intelligence of NPCs is measured by the complexity and degree of change of these behavior trees.
The problem for developers is that for every possible player action, new branches need to be added to the behavior tree to make NPCs look truly realistic. Not only that, but as more reactions increase, so does each NPC need more computing power.
EA's new patent is "readable and editable" NPC behavior. Essentially, it describes a way to dynamically edit these NPC behavior trees, rather than adding complexity to include all possible NPC actions. It also describes reading specific player data, such as location and health, and using "reinforcement learning" to adapt to NPC behavioral goals. But it's unclear whether the system will be used in future EA games.
From the description, the system can be applied to NPC interactions, especially reinforcement learning. NPCs are not only able to recognize the current player's actions, but also remember the player's previous actions. Or they can remember where they last saw the player, or how much health they have, which will create a very immersive game world.
The patent also states that the generated behavioral tree will be plotted from these reinforcement learning data. This means that two different players may interact differently with each NPC in the game, depending not only on their game choices, but also on their play style.