US 12,265,917 B2
Agent-based training of artificial intelligence character models
Ilya Gelfenbeyn, Palo Alto, CA (US); Mikhail Ermolenko, Mountain View, CA (US); and Kylan Gibbs, San Francisco, CA (US)
Assigned to Theai, Inc., Mountain View, CA (US)
Filed by Theai, Inc., Mountain View, CA (US)
Filed on Apr. 28, 2023, as Appl. No. 18/140,636.
Claims priority of provisional application 63/335,923, filed on Apr. 28, 2022.
Prior Publication US 2023/0351217 A1, Nov. 2, 2023
Int. Cl. G06N 5/022 (2023.01); A63F 13/58 (2014.01); G06N 3/006 (2023.01)
CPC G06N 5/022 (2013.01) [A63F 13/58 (2014.09); G06N 3/006 (2013.01)] 19 Claims
OG exemplary drawing
 
1. A method for training an Artificial Intelligence (AI) character model, the method being implemented with a processor of a computing platform providing the AI character model, the method comprising:
receiving, by the processor from a client-side computing device being in communication with the computing platform, input data to the AI character model;
passing, by the processor, the input data through a first plurality of heuristic machine learning models to produce intermediate outputs of the first plurality of heuristic machine learning models, the intermediate outputs defining parameters associated with the AI character model;
feeding, by the processor, the intermediate outputs to a second plurality of primary machine learning models, the second plurality of primary machine learning models being configured to select, based on the parameters associated with the AI character model, an action to be performed by the AI character model in response to the input data;
upon the selecting the action, generating, by the processor, the action to be performed by the AI character model in response to the input data;
transmitting, by the processor, the action to the client-side computing device, wherein the client-side computing device presents the action to a user;
receiving, by the processor from the client-side computing device, a message including an indication that the action is incorrect and an instruction to increase a level of a specific behavioral characteristic of a communication style of the AI character model;
receiving, by the processor, from the user and via the client-side computing device, a first portion of data to be possessed by an AI character, the AI character being generated by the AI character model and performing the action in a virtual environment;
receiving, by the processor, from the user, via the client-side computing device, a further instruction directing the AI character to forget a second portion of data, the second portion of data being stored in a knowledge store associated with the AI character model; and
based on the message, modifying, by the processor, parameters of the AI character model such that the AI character model executes a corrected action in response to the input data, wherein the parameters of the AI character model include runtime parameters generated by the processor based on interactions of the AI character model with the user and a scene surrounding the AI character model in the virtual environment, wherein the modifying the parameters of the AI character model includes:
modifying the runtime parameters by increasing the level of the specific behavioral characteristic by a predetermined percentage; and
updating the knowledge store associated with the AI character model, the knowledge store storing a plurality of data to be possessed by the AI character, the plurality of data storing the second portion of data, wherein the updating the knowledge store includes:
storing, based on the message and the instruction, the first portion of data in the knowledge store;
based on the further instruction, preventing the AI character model from using the second portion of data in generation of the corrected action; and
setting the AI character model to generate the corrected action based on the first portion of data, the corrected action being performed by the AI character in the virtual environment.