Reaction AI ReInvented

•Dynamic Non-player character Combat Artificial Intelligence using Machine Learning Multiclass prediction engine (Reaction AI + Microsoft ML.NET)

•A* pathfinding

•Automatic audio detection and following for NPC

NOTE: requires: x64 platform

Built on top of Reaction AI Steroid Edition with optimizations and now using Microsoft ML .NET Machine Learning for Multiclass predictions for NPCs' moves.

Still includes A* pathfinding as well as automatic follow sound source

see User Guide

Reaction AI is an artificial intelligence engine to be used for game developers that have nonplayer characters governed by AI or other game components that rely on artificial intelligence for reactions based on actions by the avatar. It can also be used for other systems that require learning systems and provides an intelligent output based on input. Reaction AI models the human Mind and neural networks in its design. It learns over time and uses past events in a memory system to determine the next best reaction. However, it is pseudo-dynamic in the resulting output based on how well the “mind” is functioning at that time based on stimuli. The “mind” is divided into four components, each sized proportionally to the human mind.

Once there is enough data, it will train an ML .NET multiclassification model and use that model to make predictions of what the NPC action should be. The prediction is based on a feature vector that includes providing a parameter for the desired success value, and then vending the respective action based on that with the other features provided for the training and prediction.

If a model has been trained, it will be saved locally and used for further predictions. The model can continuously get trained with additional data over time. The training data is also saved locally and can be used to train the model upon making an API call. The training data size limit is also configurable and will not exceed the amount specified (to not allow the data size to increase beyond a desired amount).

The multiclass training and predictions use a Multiclass Maximum Entropy training algorithm. Before the model is trained, the API will utilize the base Reaction AI algorithm for determining the NPC move. As mentioned, once the Machine Learning Model is trained, it will then start making prediction using the trained model. Furthermore, the predictions will also introduce a reaction using the base algorithm (not the model) once every 5 times as a means to introduce “noise” to allow the next iteration of training the model to be more dynamic.