Extensible agents in agent-based generative models
A system, method, and computer-readable medium for generating factual and/or counterfactual data are described. This may have the effect of improving the complexity of data available for training machine learning models. The models may include agent-based models (ABMs) in which the agent definitions are decoupled from the simulation. In one or more aspects, some agents may have attributes and associated behaviors that permit them to be reused in different ABMs to simulate different systems.
1. A computer-implemented method comprising:
storing, in a storage, one or more agent complex probability distribution definitions, wherein each agent complex probability distribution definition comprises a plurality of agent attribute probability distribution definitions and agent behavior probability distribution definitions;
receiving, for a first simulation, a first simulation specification, wherein the first simulation specification comprises a first list of agent complex probability distribution definitions;
generating, using a random number generator, first agent attribute probability distributions for each of the first list of agent complex probability distribution definitions;
generating, based on the first agent attribute probably distributions, the first simulation;
generating, based on the first simulation and using the random number generator, first agent behaviors for first steps of the first simulation;
outputting, based on the first steps of the first simulation, a first synthetic dataset;
receiving, for a second simulation, a second simulation specification, wherein the second simulation specification comprises a second list of agent complex probability distribution definitions;
generating, using the random number generator, second agent attribute probability distributions for each of the second list of agent complex probability distribution definitions;
generating, based on the second agent attribute probably distributions, the second simulation;
generating, based on the second simulation and using the random number generator, second agent behaviors for second steps of the second simulation;
outputting, based on the second steps of the second simulation, a second synthetic dataset; and
training, based on the second synthetic dataset, a machine-learning model,
wherein the first list of agent complex probability distribution definitions and the second list of agent complex probability distribution definitions include at least one common agent complex probability distribution definition.
2. The computer-implemented method of claim 1 ,
wherein a difference between the first simulation specification and the second simulation specification comprises, for the at least one common agent complex probability distribution definition, a different combination of agent attribute probability distributions.
3. The computer-implemented method of claim 1 ,
wherein a difference between the first simulation specification and the second simulation specification comprises, for the at least one common agent complex probability distribution definition, a different combination of agent behavior probability distributions to be simulated.
4. The computer-implemented method of claim 1 ,
wherein fields of the first synthetic dataset are different from fields of the second synthetic dataset.
5. The computer-implemented method of claim 1 ,
wherein the first agent attribute probability distributions comprise at least one agent attribute probability distribution that, after generation of the first simulation, comprises a sampled value.
6. The computer-implemented method of claim 1 ,
wherein generating the first simulation comprises sampling each agent attribute probability distribution to determine a corresponding value.
7. The computer-implemented method of claim 1 ,
wherein the first simulation specification comprises a first set of synthetic data fields to be output,
wherein the second simulation specification comprises a second set of synthetic data fields to be output, and
wherein the first set of synthetic data fields is different from the second set of synthetic data fields.
8. The computer-implemented method of claim 1 , wherein the generating the first agent behaviors for the first steps of the first simulation further comprise:
iteratively sampling first agent behavior probability distributions.
9. The computer-implemented method of claim 8 , wherein outputting the first synthetic dataset comprises:
streaming, per executed step, the first synthetic dataset.
10. The computer-implemented method of claim 8 ,
wherein the first synthetic dataset comprises synthetic data based on execution of multiple steps.
11. The computer-implemented method of claim 1 , further comprising:
receiving instructions to modify, for the first simulation specification, a quantity of instances of agent complex probability distribution definitions to be instantiated;
modifying, based on the instructions, the first simulation specification;
generating, based on the modified first simulation specification and via the random number generator, a third simulation comprising third agent attribute probability distributions;
generating, based on the third simulation and via the random number generator, third agent behaviors for third steps of the third simulation; and
outputting, based on the third simulation, a third synthetic dataset.
12. The computer-implemented method of claim 1 ,
wherein the agent complex probability distribution definitions comprise probability monads, and
wherein the probability monads are a complex probability distribution composed of agent attribute probability distributions of agent attribute probability monads.
13. The computer-implemented method of claim 1 ,
wherein generating the first simulation comprises generating a simulation monad, and
wherein the simulation monad is a complex probability distribution composed of agent behavior probability distributions of agent behavior probability monads.
14. The computer-implemented method of claim 1 ,
wherein first agent behavior probability distribution definitions describe one or more actions,
wherein one or more actions comprise action probability distributions, and
wherein the action probability distributions are a complex probability distribution composed of agent behavior probability distributions.
15. An apparatus comprising:
one or more processors; and
memory storing instructions that, when executed by the one or more processors, cause the apparatus to:
receive, for a first simulation, a first simulation specification, wherein the first simulation specification comprises a first list of agent complex probability distribution definitions, wherein the agent complex probability distribution definitions comprise a plurality of agent attribute probability distribution definitions and agent behavior probability distribution definitions;
generate, using a random number generator, first agent attribute probability distributions for each of the first list of agent complex probability distribution definitions;
generate, based on the first agent attribute probably distributions, the first simulation;
generate, based on the first simulation and using the random number generator, first agent behaviors for first steps of the first simulation;
output, based on the first steps of the first simulation, a first synthetic dataset;
receive, for a second simulation, a second simulation specification, wherein the second simulation specification comprises a second list of agent complex probability distribution definitions;
generate, using the random number generator, second agent attribute probability distributions for each of the second list of agent complex probability distribution definitions;
generate, based on the second agent attribute probably distributions, the second simulation;
generate, based on the second simulation and using the random number generator, second agent behaviors for second steps of the second simulation;
output, based on the second steps of the second simulation, a second synthetic dataset; and
train, based on the second synthetic dataset, a machine-learning model,
wherein the first list of agent complex probability distribution definitions and the second list of agent complex probability distribution definitions include at least one common agent complex probability distribution definition.
16. The apparatus of claim 15 ,
wherein a difference between the first simulation specification and the second simulation specification comprises, for the at least one common agent complex probability distribution definition, a different combination of agent attribute probability distributions.
17. The apparatus of claim 15 ,
wherein a difference between the first simulation specification and the second simulation specification comprises, for the at least one common agent complex probability distribution definition, a different combination of agent behavior probability distributions to be simulated.
18. The apparatus of claim 15 ,
wherein fields of the first synthetic dataset are different from fields of the second synthetic dataset.
19. The apparatus of claim 15 ,
wherein the first agent attribute probability distributions comprise at least one agent attribute probability distribution that, after generation of the first simulation, comprises a sampled value.
20. One or more non-transitory media storing instructions that, when executed by one or more processors, cause the one or more processors to perform steps comprising:
storing one or more agent complex probability distribution definitions, wherein each agent complex probability distribution definition comprises a plurality of agent attribute probability distribution definitions and agent behavior probability distribution definitions;
receiving, for a first simulation, a first simulation specification, wherein the first simulation specification comprises a first list of agent complex probability distribution definitions;
causing display of a graphical interface of the first simulation specification, wherein the graphical interface is configured to display one or more agent complex probability distribution definitions;
receiving user interactions with the graphical interface, wherein the user interactions are to modify a specific agent attribute probability distribution definition of the one or more agent complex probability distribution definitions;
modifying, based on the received user interactions, the one or more agent complex probability distribution definitions in the first simulation specification;
generating, using a random number generator, first agent attribute probability distributions for each of the first list of agent complex probability distribution definitions including the one or more modified agent complex probability distribution definitions;
generating, based on the first agent attribute probably distributions, the first simulation;
generating, based on the first simulation and using the random number generator, first agent behaviors for first steps of the first simulation;
outputting, based on the first steps of the first simulation, a first synthetic dataset;
receiving, for a second simulation, a second simulation specification, wherein the second simulation specification comprises a second list of agent complex probability distribution definitions;
generating, using the random number generator, second agent attribute probability distributions for each of the second list of agent complex probability distribution definitions;
generating, based on the second agent attribute probably distributions, the second simulation;
generating, based on the second simulation and using the random number generator, second agent behaviors for second steps of the second simulation;
outputting, based on the second steps of the second simulation, a second synthetic dataset; and
training, based on the second synthetic dataset, a machine-learning model,
wherein the first list of agent complex probability distribution definitions and the second list of agent complex probability distribution definitions include at least one common agent complex probability distribution definition.