Invented by Breitweiser; Edward W., State Farm Mutual Automobile Insurance Company

In today’s world, computer models learn from real data, but getting enough of the right data is hard and slow. This patent application shows a way to use pretend (simulated) responses and real user feedback together to teach computer models how real people or groups might respond to questions. Let’s break down how this works, why it matters, and what’s new about it.

Background and Market Context

Companies, schools, and other groups make big decisions every day. They change prices, create new ads, launch new products, and even adjust company rules. Before they do this, they want to know: “How will our people react?” But getting answers from every real person in a group is impossible. Surveys take time and might not give the full picture.

Businesses use computer models to guess what people will say or do. These models use data from the past, or from big collections of public data. But public data is not always a good match for a certain group, like a company’s customers, a market segment, or a team’s fans. If a business wants to know what their own customers, not just the public, would think about a new feature, they may not have enough special data to train their model.

This is a big problem. Training a new model from scratch every time the target group changes takes too long. It costs a lot, and sometimes the data about the group just isn’t there or is too private to use. Companies need a way to quickly get good, group-specific answers without waiting for months of data collection.

The solution offered in this patent is to make “pretend” (simulated) people, called simulated characters, who stand in for real group members. The system then creates a pretend population made from these simulated characters. It can answer questions on their behalf, and then real users (like company staff or customers) give feedback on those answers. The feedback helps the computer model keep learning and get better over time. This way, companies get quick, tailored insights for their own groups without waiting for real-world data to arrive.

Scientific Rationale and Prior Art

Traditional computer models learn by looking at lots of real-life answers and actions. For example, if a company wants to teach a model to predict how people will respond to a new price, it needs past records of similar price changes and how people reacted. These models use methods like neural networks, decision trees, or even simple rules learned from data.

But these models usually use data from the general public, not from a specific group. For example, a chatbot might be trained on millions of public conversations, but if you want it to act like your company’s customers, you need data just from them. That’s not easy. You might not have enough, or it might take months to collect. Also, if the group changes—say, you want to model teenage customers instead of retirees—you have to start over.

Some older methods try to build customer segments or user personas, but these are often static and not very flexible. If you want to ask “What would a 30-year-old mom from Texas who shops online say about this?” you might not have enough real people in that category to learn from.

In the past, companies have tried to make “synthetic data” or “virtual agents” to fill in the gaps. But these fake people were often simple copies, without enough variation to match real group diversity. They also didn’t learn from new user feedback—they stayed the same unless rebuilt from scratch.

This patent stands out because it combines simulated characters with real user feedback in a loop. The computer system first builds a pretend population that reflects the group you care about, based on known traits (like age, gender, or shopping habits). It then answers your questions using this pretend group. After that, real users review the answers and give feedback: did the answers seem right or wrong? The system uses these signals to keep retraining its models—so it gets smarter and more accurate for your specific needs over time.

This is different from old approaches where models only trained on static data, or only on public data, or only on feedback from general users. Here, you get group-specific, dynamic training that adapts as your group or your questions change.

Invention Description and Key Innovations

Let’s walk through how the invention works, step by step. The whole process is done by a computer system running special software. The system can work on servers, phones, laptops, or wearables.

First, a user (like a company manager) tells the system some facts about their actual group. These are called known-member characteristics. For instance, they might say, “My target group is women, ages 25-35, living in California, who bought from us in the last year.” They might also ask a question like, “If we offer free shipping, will more people buy?”

The system uses these facts to build a “simulated population.” It creates pretend people—simulated characters—who match the traits of the real group. Each pretend person has values for things like age, location, buying habits, and so on. The system can make these characters by pulling from a database of past real people, or by using computer models to generate new, similar ones.

The next step is to match each real group member (if known) to a simulated character. For example, if you know Jane is a 32-year-old mother who shops online, the system finds or creates a simulated character with similar traits.

Once the pretend group is set, the system can answer questions on their behalf. It uses another model—a response-generating model—to predict how each simulated character would answer the user’s question. For example, it might predict that most young moms would like free shipping, but a few might not care.

The system shows these simulated answers on a user interface. Now, the real user (say, a manager or a marketer) reads the answers and gives feedback. Maybe they think the answers are realistic, or maybe they notice something is off—like, “No, our customers would care more about fast delivery than free shipping.”

This feedback is gold. The system takes it and uses it to retrain its models. That means the next time it builds a simulated population or predicts answers, it does a better job. Over time, with each cycle of feedback, the system becomes more and more tuned to the real group’s needs and opinions.

There are even more smart features in the design:

– The system can use different sources to build simulated characters. It can pull real data, create new characters with machine learning, or blend both. This lets it adapt to situations where little or no real data exists.

– It handles variations. For example, if you know your group is mostly women aged 30, but you’re not sure about their income or hobbies, the system can create simulated characters with a range of possible values, based on statistics or expert input.

– It can train multiple models at once. There’s a model for making characters, a model for building the population, and a model for predicting answers. Each can learn from feedback and keep improving.

– It works for many types of groups: customers of a store, owners of a certain car, members of a sports team’s fan club, or any group you can describe with traits.

– The system can run on many types of devices, from large servers to smartwatches or phones, making it easy to use anywhere.

– User feedback can come in many forms: a thumbs-up or thumbs-down, a rating, or even real-world actions (like whether people actually buy after a price change).

– The design supports privacy. If you can’t use real customer data, you can still build simulated characters based on general traits, without exposing anyone’s private details.

This approach is flexible and powerful. It lets users test ideas fast—like a new marketing plan, a product tweak, or a company policy—and see how their special group might react, all before taking action in the real world. If the simulated answers seem off, user feedback tunes the models, making them smarter for the next round of questions.

Conclusion

This patent application brings a fresh way to predict how real groups will respond to changes, without waiting for months of real data. By using pretend people (simulated characters) and learning from real user feedback, the system gets smarter and more accurate for any group you care about. The invention saves time, protects privacy, and adapts quickly to new questions or changing groups. This is a big step forward for companies, organizations, and anyone who wants fast, group-specific insights in a changing world.

Click here https://ppubs.uspto.gov/pubwebapp/ and search 20250217630.