Granted Patent B1

US 12,210,849 · App. 18/623,959 · Granted Jan 28, 2025

Inventor: Kino High Coursey (Colleyville, TX)

G06F40/58G06F16/90335G06F40/284G06F40/35

Abstract

Claims (79)

1. An Artificial Intelligence (AI) language virtual agent having self-improvement features which incorporate language modeling and tree search, the virtual agent comprising:

a data processing system having program instructions implementing the virtual agent;

the implemented virtual agent accepting as input a current situational description that includes natural language input received from an external agent, properties regarding the qualities of the virtual agent, and indicia regarding subject matter context of a present conversation;

wherein the qualities of the virtual agent include temperament and textual tendencies;

wherein the indicia regarding subject matter context includes textual logs from recent conversational exchanges;

wherein the current situational description includes audio, visual, and tactile inputs from the external agent which are collected proximate to the virtual agent;

wherein the virtual agent employs reinforcement learning;

a database of one or more language models, conversation logs storing text from prior textual exchanges, and reference conversations for training;

a combination of self-play engines that train on the language models with self-play and external interaction engines for communicating with external agents; and

a combination of self-moving modules that advance the method of the external agent communicating with the virtual agent via a combination of textual exchanges and one or more audio, visual, and tactile inputs into the virtual agent; and

wherein the virtual agent utilizing tree search techniques in combination with the one or more language models for outputting optimized responses to the current situation description, and wherein the virtual agent responds with self-optimized textual expression to verbal input in combination with the audio, visual, tactile, and other sensory inputs.

2. The AI language virtual agent according to claim 1 , wherein the virtual agent is further configured to:

select an optimal response to a current situation description with respect to immediate and non-immediate goals and outcomes by using one or more language models to form the optimal response search process;

wherein the language models are used to anticipate future outcomes and future responses;

wherein the language models are also used to evaluate the desirability of projected future outcomes;

convert the language models evaluation into a format suitable to guide tree search methods; and

wherein information regarding the desired goals or outcomes may be included in the current situational description.

3. The AI language virtual agent according to claim 1 , further comprising a feedback mechanism that allows the external agent to rate and review the performance of the responses.

4. The AI language virtual agent according to claim 1 , further comprising a dedicated emotion recognition module using multimodal emotion recognition.

5. The AI language virtual agent according to claim 1 , further comprising a module for understanding visual and auditory metaphors, sarcasms, and idiomatic expressions.

6. The AI language virtual agent according to claim 1 , wherein the virtual agent is further configured to:

utilize one or more language models to generate evaluative statements or annotations on the desirability of a situational description;

translate symbolic evaluative statements into numeric or vector format suitable for use in common state space search systems;

select an initial action from an initial set based on the outcome of the complete search process;

perform an action or set of actions based on the initial action selected;

record the system and environmental interactions to one or more logs;

provide for the continuing training of the language models used by the virtual agent using reference material and direct experience; and

provide self-improvement of the entire process through iterative retraining of the language models, and through use of self-play to explore the space of novel possible situations and to discover paths towards positive resolutions in those situations.

7. The AI virtual agent according to claim 1 , further comprising a learning module for understanding and utilizing emerging online languages, lingoes, and slangs that evolve continuously.

8. An Artificial Intelligence (AI) language virtual agent having self-improvement features which incorporate language modeling and tree search techniques, the AI language virtual agent comprising:

a data processing system having program instructions implementing the AI language virtual agent;

sensory inputs having one or more audio, visual, and tactile receptors located proximate to the virtual agent for collecting indicia of verbal and situational indexes, wherein said sensory inputs are connected to said data processing system for inputting the indicia of verbal and situational indexes;

said data processing system processing said indicia of verbal and situational indexes to define a current situational description;

said current situational description including natural language input, qualities of the virtual agent, and indicia regarding subject matter context of a present conversation, wherein said qualities of the virtual agent include temperament and textual tendencies, and the indicia regarding subject matter context includes textual logs from recent conversational exchanges;

wherein the virtual agent employs reinforcement learning;

a database adapted for access by said data processing system and having one or more language models, conversation logs storing text from prior textual exchanges, and reference conversations for training;

one or more self-play engines for training said one or more language models with self-play; and

wherein the AI virtual agent utilizes tree search techniques in combination with said one or more language models for outputting textual responses to the current situation description.

9. The AI language virtual agent according to claim 8 ,

wherein the virtual agent utilizes Monte Carlo Tree Search (MCTS) techniques in combination with the one or more language models for outputting textual responses to the current situation description;

wherein the virtual agent responds with textual expression to verbal input in combination with the audio, visual, tactile, and other sensory inputs after merging an analysis of basic sentiment analysis and goal-oriented analysis of the AI virtual agent; and

wherein the sentiment analysis is determined by the text being tokenized into a list of words and phrases, a sentiment value is assigned for each token in which a summation of the sentiment values are computed and then normalized as compared to an expected value to determine respective normalized scores, and then the respective normalized scores are compared to one or more indexes which are summed to provide discrete output values which provide a respective sentiment value for the text being analyzed;

wherein the virtual agent may engage with proof of work system.

10. The AI language virtual agent according to claim 8 , wherein the virtual agent is further configured to:

select an optimal response to a current situation description with respect to immediate and non-immediate goals or outcomes by using one or more language models to form the optimal response search process;

said virtual agent further comprising a contextual interface interchangeable with an external state simulator or server that maintains real-time tracking of global states, and wherein the interface accommodates natural language instructions or action mandates from the AI virtual agent and reciprocates with natural language descriptions of the current state.

11. The AI language virtual agent according to claim 10 , wherein the external state simulator or server provides comprehensive account keeping details in addition to producing numeric evaluations of the state, enhancing the agent's understanding of the surrounding environment.

12. The AI language virtual agent according to claim 8 , wherein the virtual agent incorporates LLM capable of producing executable code given a problem description, thereby providing a viable solution to the specified problem.

13. The AI language virtual agent according to claim 8 , wherein the LLM drives a dynamic state simulator by generating executable code that alters one state representation into another, demonstrating a broader state expansion and efficient leaf evaluation.

14. The AI language virtual agent according to claim 8 , wherein the virtual agent utilizes any available resources, planning, or computing capabilities by virtue of the code produced for state expansion and leaf evaluation to provide optimum responses.

15. The AI language virtual agent according to claim 8 , wherein the output generated through the dynamic state simulation driven by LLM-written code is translatable into natural language or various compatible embeddings for improved interaction with external agents.

16. An Artificial Intelligence (AI) language virtual agent having self-improvement features which incorporate language modeling and tree search techniques, the AI language virtual agent comprising:

a data processing system having program instructions implementing the AI language virtual agent;

sensory inputs having one or more audio, visual, and tactile receptors located proximate to the virtual agent for collecting indicia of verbal and situational indexes, wherein said sensory inputs are connected to said data processing system for inputting the indicia of verbal and situational indexes;

wherein said data processing system processing said indicia of verbal and situational indexes to define a current situational description;

said current situational description including natural language input, qualities of the virtual agent, and indicia regarding subject matter context of a present conversation, wherein said qualities of the virtual agent include temperament and textual tendencies, and the indicia regarding subject matter context includes textual logs from recent conversational exchanges;

a database adapted for access by said data processing system and having one or more language models, conversation logs storing text from prior textual exchanges, and reference conversations for training;

one or more self-play engines training said one or more language models with self-play; and

wherein the AI virtual agent utilizes Monte Carlo Tree Search (MCTS) techniques in combination with said one or more language models for outputting textual responses to the current situation description; and

wherein the AI language virtual agent employs the language models for latent space exploration and optimization by transforming verbal and situational indicia into a lower-dimensional or alternate representation that uncovers the underlying semantics, making the state search process more efficient and accurate.

17. The AI language virtual agent according to claim 16 , wherein the virtual agent is further configured to:

select an optimal response to a current situation description with respect to immediate and non-immediate goals and outcomes by using one or more language models to form the optimal response search process;

wherein the language models are used to anticipate future outcomes and future responses;

wherein the language models are also used to evaluate the desirability of projected future outcomes;

convert the language models evaluation into a format suitable to guide tree search methods;

wherein information regarding the desired goals or outcomes may be included in the current situational description;

wherein the virtual agent utilizes MCTS techniques in combination with the one or more language models for outputting textual responses to the current situation description;

wherein the virtual agent responds with textual expression to verbal input in combination with the audio, visual, tactile, and other sensory inputs after merging an analysis of basic sentiment analysis and goal oriented analysis of the AI virtual agent; and

wherein the sentiment analysis is determined by the text being tokenized into a list of words and phrases, a sentiment value is assigned for teach token in which a summation of the sentiment values are computed and then normalized as compared to an expected value to determine respective normalized scores, and then the respective normalized scores are compared to one or more indexes which are summed to provide discrete output values which provide a respective sentiment value for the text being analyzed; and

wherein the virtual agent employs a dedicated emotion recognition module using multimodal emotion recognition techniques for processing sensory inputs and improving sentiment analysis.

18. The AI language virtual agent according to claim 16 , wherein the virtual agent is further configured to:

construct the contextual descriptions to maximally represent the current situation while meeting any processing constraints of the language models used;

provide nodes of the states of the search space which contain contextual descriptions and search control information;

receive an initial starting contextual description and search control information; and

explore the state of possible future contexts by using one or more language models provided with sufficient information in the form of the contextual descriptions and prompts that cause the generation of new situation descriptions;

wherein a blockchain-based secure database may be utilized for storing and accessing the data for training and self-play in a secure, transparent and immutable manner.

19. The AI language virtual agent according to claim 16 , further comprising:

a contextual interface interchangeable with an external state simulator or server that maintains real-time tracking of global states, and wherein the interface accommodates natural language instructions or action mandates from the AI agent and reciprocates with natural language descriptions of the current state.

20. The AI language virtual agent according to claim 19 , wherein the external state simulator or server provides comprehensive account keeping details in addition to producing numeric evaluations of the state, enhancing the agent's understanding of the surrounding environment and its responses to the same.

Continuity (3)

References Cited (199)

Cited By (5)

US 12,505,137 US 12,561,533 US 12,608,582 US 12,645,831 US 12,694,048