Granted Patent B2

US 12,373,705 · App. 17/433,724 · Granted Jul 29, 2025

Inventors: Theodore David Harris (San Francisco, CA); Tatiana Korolevskaya (Mountain View, CA); Yue Li (Sunnyvale, CA); Craig O'Connell (Foster City, CA)

Assignee: Visa International Service Association

G06N3/126G06N5/04

Abstract

Claims (185)

1. A method for controlling genetic learning for predictive models using predefined strategies, comprising:

loading, with at least one processor, a master template comprising a number of agents of a plurality of agents, wherein the master template further comprises goal data associated with at least one condition;

for each agent of the plurality of agents:

selecting, with at least one processor, a type of predictive model from a plurality of types of predictive models as an agent predictive model;

selecting, with at least one processor, a strategy from a plurality of predefined strategies as an agent strategy;

generating, with at least one processor, a plurality of candidate genomes, each candidate genome comprising a plurality of genes, each gene associated with a respective feature of the agent predictive model;

determining, with at least one processor, a fit of each candidate genome of the plurality of candidate genomes to the agent strategy; and

selecting, with at least one processor, a candidate genome from the plurality of candidate genomes as the agent genome based on the fit of the candidate genome;

for each epoch of a plurality of epochs:

for each agent of the plurality of agents:

performing, with at least one processor, a plurality of training iterations based on the agent predictive model and the agent genome; and

determining, with at least one processor, a fitness of the agent predictive model after the plurality of training iterations;

determining, with at least one processor, a first subset of agents with a highest fitness;

for each agent of the first subset of agents:

generating, with at least one processor, at least one new agent to add to the plurality of agents; and

merging, with at least one processor, a genome of the at least one new agent with the agent genome of at least one agent of the first subset of agents; and

determining, with at least one processor, whether the at least one agent predictive model of the plurality of agents satisfies the at least one condition associated with the goal data; and

terminating, with at least one processor, the plurality of epochs based on determining the at least one agent predictive model of the plurality of agents satisfies the at least one condition associated with the goal data.

2. The method of claim 1 , wherein the master template further comprises a plurality of agent templates, wherein selecting the type of predictive model comprises selecting, with at least one processor, the type of predictive model based on a subset of the plurality of types of predictive models associated with a respective agent template of the plurality of agent templates for each agent.

3. The method of claim 1 , further comprising, for each epoch of the plurality of epochs:

determining, with at least one processor, a second subset of agents with a lowest fitness.

4. The method of claim 3 , further comprising, for each epoch of the plurality of epochs:

filtering, with at least one processor, the at least one new agent based on the second subset of agents.

5. The method of claim 1 , wherein generating the at least one new agent comprises generating, with at least one processor, at least three new agents.

6. A method for controlling genetic learning for predictive models using predefined strategies, comprising:

loading, with at least one processor, a master template comprising a number of agents of a plurality of agents, wherein the master template further comprises goal data associated with at least one condition;

for each agent of the plurality of agents:

selecting, with at least one processor, a type of predictive model from a plurality of types of predictive models as an agent predictive model;

selecting, with at least one processor, a strategy from a plurality of predefined strategies as an agent strategy;

generating, with at least one processor, a plurality of candidate genomes, each candidate genome comprising a plurality of genes, each gene associated with a respective feature of the agent predictive model;

determining, with at least one processor, a fit of each candidate genome of the plurality of candidate genomes to the agent strategy; and

selecting, with at least one processor, a candidate genome from the plurality of candidate genomes as the agent genome based on the fit of the candidate genome;

for each epoch of a plurality of epochs:

for each agent of the plurality of agents:

performing, with at least one processor, a plurality of training iterations based on the agent predictive model and the agent genome; and

determining, with at least one processor, a fitness of the agent predictive model after the plurality of training iterations;

determining, with at least one processor, a first subset of agents with a highest fitness;

for each agent of the first subset of agents:

generating, with at least one processor, at least one new agent to add to the plurality of agents; and

merging, with at least one processor, a genome of the at least one new agent with the agent genome of at least one agent of the first subset of agents,

wherein the master template further comprises a plurality of agent templates, wherein selecting the type of predictive model comprises selecting, with at least one processor, the type of predictive model based on a subset of the plurality of types of predictive models associated with a respective agent template of the plurality of agent templates for each agent,

wherein each agent template of the plurality of agent templates further comprises at least one predictive model template, wherein selecting the type of predictive model comprises selecting, with at least one processor, one of the at least one predictive model template based on the respective agent template for each agent.

7. A method for controlling genetic learning for predictive models using predefined strategies, comprising:

for each agent of a plurality of agents:

selecting, with at least one processor, a type of predictive model from a plurality of types of predictive models as an agent predictive model;

selecting, with at least one processor, a strategy from a plurality of predefined strategies as an agent strategy;

generating, with at least one processor, a plurality of candidate genomes, each candidate genome comprising a plurality of genes, each gene associated with a respective feature of the agent predictive model;

determining, with at least one processor, a fit of each candidate genome of the plurality of candidate genomes to the agent strategy; and

selecting, with at least one processor, a candidate genome from the plurality of candidate genomes as the agent genome based on the fit of the candidate genome;

for each epoch of a plurality of epochs:

for each agent of the plurality of agents:

performing, with at least one processor, a plurality of training iterations based on the agent predictive model and the agent genome; and

determining, with at least one processor, a fitness of the agent predictive model after the plurality of training iterations;

determining, with at least one processor, a first subset of agents with a highest fitness;

for each agent of the first subset of agents:

generating, with at least one processor, at least one new agent to add to the plurality of agents; and

merging, with at least one processor, a genome of the at least one new agent with the agent genome of at least one agent of the first subset of agents,

wherein generating the at least one new agent comprises generating, with at least one processor, at least three new agents,

wherein the at least three new agents comprise a first new agent, a second new agent, and a third new agent, and wherein merging the genome of the at least one new agent with the agent genome of at least one agent of the first subset of agents comprises:

merging, with at least one processor, the genome of the first new agent with the agent genome of a first agent of the first subset of agents having a most different strategy;

merging, with at least one processor, the genome of the second new agent with the agent genome of a second agent of the first subset of agents having a highest fitness; and

merging, with at least one processor, the genome of the third new agent with the agent genome of a third agent of the first subset of agents having a most similar winning strategy.

8. A system for controlling genetic learning for predictive models using predefined strategies, comprising:

at least one processor; and

at least one non-transitory computer-readable medium including one or more instructions that, when executed by at least one processor, cause the at least one processor to:

load a master template comprising a number of agents of a plurality of agents, wherein the master template further comprises goal data associated with at least one condition;

for each agent of the plurality of agents:

select a type of predictive model from a plurality of types of predictive models as an agent predictive model;

select a strategy from a plurality of predefined strategies as an agent strategy;

generate a plurality of candidate genomes, each genome comprising a plurality of genes, each gene associated with a respective feature of the agent predictive model;

determine a fit of each candidate genome of the plurality of candidate genomes to the agent strategy; and

select a candidate genome from the plurality of candidate genomes as the agent genome based on the fit of the candidate genome;

for each epoch of a plurality of epochs:

for each agent of the plurality of agents:

perform a plurality of training iterations based on the agent predictive model and the agent genome; and

determine a fitness of the agent predictive model after the plurality of training iterations;

determine a first subset of agents with a highest fitness;

for each agent of the first subset of agents:

generate at least one new agent to add to the plurality of agents; and

merge a genome of the at least one new agent with the agent genome of at least one agent of the first subset of agents; and

determine whether the at least one agent predictive model of the plurality of agents satisfies the at least one condition associated with the goal data; and

terminate the plurality of epochs based on determining the at least one agent predictive model of the plurality of agents satisfies the at least one condition associated with the goal data.

9. The system of claim 8 , wherein the master template further comprises a plurality of agent templates, wherein selecting the type of predictive model comprises selecting the type of predictive model based on a subset of the plurality of types of predictive models associated with a respective agent template of the plurality of agent templates for each agent.

10. The system of claim 8 , wherein the one or more instructions, when executed by the at least one processor, further cause the at least one processor to, for each epoch of the plurality of epochs:

determine a second subset of agents with a lowest fitness.

11. The system of claim 10 , wherein the one or more instructions, when executed by the at least one processor, further cause the at least one processor to, for each epoch of the plurality of epochs:

filter the at least one new agent based on the second subset of agents.

12. The system of claim 8 , wherein generating the at least one new agent comprises generating at least three new agents.

13. A system for controlling genetic learning for predictive models using predefined strategies, comprising:

at least one processor; and

at least one non-transitory computer-readable medium including one or more instructions that, when executed by at least one processor, cause the at least one processor to:

load a master template comprising a number of agents of a plurality of agents;

for each agent of the plurality of agents:

select a type of predictive model from a plurality of types of predictive models as an agent predictive model;

select a strategy from a plurality of predefined strategies as an agent strategy;

generate a plurality of candidate genomes, each genome comprising a plurality of genes, each gene associated with a respective feature of the agent predictive model;

determine a fit of each candidate genome of the plurality of candidate genomes to the agent strategy; and

select a candidate genome from the plurality of candidate genomes as the agent genome based on the fit of the candidate genome;

for each epoch of a plurality of epochs:

for each agent of the plurality of agents:

perform a plurality of training iterations based on the agent predictive model and the agent genome; and

determine a fitness of the agent predictive model after the plurality of training iterations;

determine a first subset of agents with a highest fitness;

for each agent of the first subset of agents:

generate at least one new agent to add to the plurality of agents; and

merge a genome of the at least one new agent with the agent genome of at least one agent of the first subset of agents,

wherein the master template further comprises a plurality of agent templates, wherein selecting the type of predictive model comprises selecting the type of predictive model based on a subset of the plurality of types of predictive models associated with a respective agent template of the plurality of agent templates for each agent,

wherein each agent template of the plurality of agent templates further comprises at least one predictive model template, wherein selecting the type of predictive model comprises selecting one of the at least one predictive model template based on the respective agent template for each agent.

14. A system for controlling genetic learning for predictive models using predefined strategies, comprising:

at least one processor; and

at least one non-transitory computer-readable medium including one or more instructions that, when executed by at least one processor, cause the at least one processor to:

for each agent of a plurality of agents:

select a type of predictive model from a plurality of types of predictive models as an agent predictive model;

select a strategy from a plurality of predefined strategies as an agent strategy;

generate a plurality of candidate genomes, each genome comprising a plurality of genes, each gene associated with a respective feature of the agent predictive model;

determine a fit of each candidate genome of the plurality of candidate genomes to the agent strategy; and

select a candidate genome from the plurality of candidate genomes as the agent genome based on the fit of the candidate genome;

for each epoch of a plurality of epochs:

for each agent of the plurality of agents:

perform a plurality of training iterations based on the agent predictive model and the agent genome; and

determine a fitness of the agent predictive model after the plurality of training iterations;

determine a first subset of agents with a highest fitness;

for each agent of the first subset of agents:

generate at least one new agent to add to the plurality of agents; and

merge a genome of the at least one new agent with the agent genome of at least one agent of the first subset of agents,

wherein generating the at least one new agent comprises generating at least three new agents,

wherein the at least three new agents comprise a first new agent, a second new agent, and a third new agent, and wherein merging the genome of the at least one new agent with the agent genome of at least one agent of the first subset of agents comprises:

merging the genome of the first new agent with the agent genome of a first agent of the first subset of agents having a most different strategy;

merging the genome of the second new agent with the agent genome of a second agent of the first subset of agents having a highest fitness; and

merging the genome of the third new agent with the agent genome of a third agent of the first subset of agents having a most similar winning strategy.

15. A computer program product for controlling genetic learning for predictive models using predefined strategies, the computer program product comprising at least one non-transitory computer-readable medium including one or more instructions that, when executed by at least one processor, cause the at least one processor to:

load a master template comprising a number of agents of a plurality of agents, wherein the master template further comprises goal data associated with at least one condition;

for each agent of the plurality of agents:

select a type of predictive model from a plurality of types of predictive models as an agent predictive model;

select a strategy from a plurality of predefined strategies as an agent strategy;

generate a plurality of candidate genomes, each genome comprising a plurality of genes, each gene associated with a respective feature of the agent predictive model;

determine a fit of each candidate genome of the plurality of candidate genomes to the agent strategy; and

select a candidate genome from the plurality of candidate genomes as the agent genome based on the fit of the candidate genome;

for each epoch of a plurality of epochs:

for each agent of the plurality of agents:

perform a plurality of training iterations based on the agent predictive model and the agent genome; and

determine a fitness of the agent predictive model after the plurality of training iterations;

determine a first subset of agents with a highest fitness;

for each agent of the first subset of agents:

generate at least one new agent to add to the plurality of agents; and

merge a genome of the at least one new agent with the agent genome of at least one agent of the first subset of agents; and

determine whether the at least one agent predictive model of the plurality of agents satisfies the at least one condition associated with the goal data; and

terminate the plurality of epochs based on determining the at least one agent predictive model of the plurality of agents satisfies the at least one condition associated with the goal data.

16. A computer program product for controlling genetic learning for predictive models using predefined strategies, the computer program product comprising at least one non-transitory computer-readable medium including one or more instructions that, when executed by at least one processor, cause the at least one processor to:

for each agent of a plurality of agents:

select a type of predictive model from a plurality of types of predictive models as an agent predictive model;

select a strategy from a plurality of predefined strategies as an agent strategy;

generate a plurality of candidate genomes, each genome comprising a plurality of genes, each gene associated with a respective feature of the agent predictive model;

determine a fit of each candidate genome of the plurality of candidate genomes to the agent strategy; and

select a candidate genome from the plurality of candidate genomes as the agent genome based on the fit of the candidate genome;

for each epoch of a plurality of epochs:

for each agent of the plurality of agents:

perform a plurality of training iterations based on the agent predictive model and the agent genome; and

determine a fitness of the agent predictive model after the plurality of training iterations;

determine a first subset of agents with a highest fitness;

for each agent of the first subset of agents:

generate at least one new agent to add to the plurality of agents; and

merge a genome of the at least one new agent with the agent genome of at least one agent of the first subset of agents,

wherein generating the at least one new agent comprises generating at least three new agents comprising a first new agent, a second new agent, and a third new agent, and wherein merging the genome of the at least one new agent with the agent genome of at least one agent of the first subset of agents comprises:

merging the genome of the first new agent with the agent genome of a first agent of the first subset of agents having a most different strategy;

merging the genome of the second new agent with the agent genome of a second agent of the first subset of agents having a highest fitness; and

merging the genome of the third new agent with the agent genome of a third agent of the first subset of agents having a most similar winning strategy.

17. A computer program product for controlling genetic learning for predictive models using predefined strategies, the computer program product comprising at least one non-transitory computer-readable medium including one or more instructions that, when executed by at least one processor, cause the at least one processor to:

load a master template comprising a number of agents of a plurality of agents, wherein the master template further comprises goal data associated with at least one condition;

for each agent of the plurality of agents:

select a type of predictive model from a plurality of types of predictive models as an agent predictive model;

select a strategy from a plurality of predefined strategies as an agent strategy;

generate a plurality of candidate genomes, each genome comprising a plurality of genes, each gene associated with a respective feature of the agent predictive model;

determine a fit of each candidate genome of the plurality of candidate genomes to the agent strategy; and

select a candidate genome from the plurality of candidate genomes as the agent genome based on the fit of the candidate genome;

for each epoch of a plurality of epochs:

for each agent of the plurality of agents:

perform a plurality of training iterations based on the agent predictive model and the agent genome; and

determine a fitness of the agent predictive model after the plurality of training iterations;

determine a first subset of agents with a highest fitness;

for each agent of the first subset of agents:

generate at least one new agent to add to the plurality of agents; and

merge a genome of the at least one new agent with the agent genome of at least one agent of the first subset of agents,

wherein the master template further comprises a plurality of agent templates, wherein selecting the type of predictive model comprises selecting the type of predictive model based on a subset of the plurality of types of predictive models associated with a respective agent template of the plurality of agent templates for each agent,

wherein each agent template of the plurality of agent templates further comprises at least one predictive model template, wherein selecting the type of predictive model comprises selecting one of the at least one predictive model template based on the respective agent template for each agent.

Assignments (1)

Continuity (1)

References Cited (16)