Tech Blog | Cofounder AI

Mar 18, 202211 min read

Optimizing Outcomes in People’s Lives: a Mathematical Approach

People have always been pushed to create something new and use innovation to solve large complex problems. However, entropy is a natural decay of what we build and often increases the level of complexity. People have to put effort into keeping what they have and put even more effort into what they desire. An example of this — is waking up, driving to work and doing the company tasks to receive pay. Or spending time with loved ones to keep the relationship fresh. The purpose of this article is to discuss a possible way to optimize people’s outcomes in the different domains of life using mathematics and machine learning.

I break this article down into two parts: (1) understanding human preferences and (2) Calculating the outcome based on Work vs Entropy.

1. Human Behavior and Preferences

Individual people rely on their instinct to make decisions. But inherently there is a deeply rooted optimization function that uses all of their knowledge (hyper graph) to derive a decision.

Cognitive Dissonance vs. Cognitive Ease

To help understand the human optimization function, we need to talk about two (2) terms. People have embedded functions called cognitive dissonance and cognitive ease. The first function focuses on your experience that contradicts your belief system while the other function makes you feel comfortable, respectively. The reason why you feel uncomfortable is because there is something that you experience that is not true or inconsistent with your current knowledge base. Inversely, if you hear something that jives with what you believe, then you will experience an easiness. Why is this important? People may have goals but if their mental state is blocking the appropriate decision making, it is unlikely they will achieve their objectives.

Note: the cognitive dissonance/ease theory is different than feeling negative or positive. I would say cognitive dissonance/ease is a derivative of how one feels. That’s why social media employs tools like this to to keep their users on their platform despite if it makes them feel good or sad; it’s about what the user believes. It’s like the movie Inception. The user has to be in cognitive ease in order to receive information or it won’t work. Cognitive dissonance will most certainly discard any information that the user receives. Why is Cognitive ease important? Even if we provide the most important aspects that can help people, they will not care unless we show tangible ways the solutions can help them enjoy their lives. In other words, show a picture of things they believe and care about and how these decisions will help them live that life.

Forming Biased Ideas and Preferences

We use learned functions that have already been optimized over time to make decisions each day — — from what we eat, to what we wear to how we interact with loves and colleagues. These models of the world or functions are just biases with a likelihood predictor given different life scenarios. Our biases are based on what we have experienced and accepted in life. As children, you mainly accept information because there is nothing to counter it. So it’s important to protect children’s minds. When people mature enough to form hundreds of millions of biases interacting with each other, they build complex understanding and their preferences are formed. Building these biases take physical work. It’s like bitcoin mining. Why does this matter? Making the connection between ones current mental state and a potential outcome is difficult. It means biases will have to change to yield the desired outcome.

A.I. and Human Preferences

Stuart Russel, a computer scientist from Berkeley notes that the way we do A.I. today will lead to issues in the future. Think about the paper clip issue. The A.I. will maximize its own goals of maximizing paper clips; thus, turning the world into a paper clip factory. Russel proposes to have A.I. seek the preferences of people. Wolchover writes on behalf of Russel’s ideas, “Instead of machines pursuing goals of their own, the new thinking goes, they should seek to satisfy human preferences; their only goal should be to learn more about what our preferences are.” [3]. This is important because, we want to bridge the gap by creating systems that care about what people believe and prefer.

“Instead of machines pursuing goals of their own, the new thinking goes, they should seek to satisfy human preferences; their only goal should be to learn more about what our preferences are.” — Wolchover [Russel]

2. Life of Work vs. Life of Entropy

To make money, keep a good marriage, keep a car from breaking down, there is a level of effort that is put into keeping these things in a good state. Work is the energy put into a subject or object to create order. Sometimes you will have to create new stuff [Elon Musk] and discard the old. For example in five years part by part you replace every piece of a marine ship. Is it still the same ship? Basically there is a level of work necessary to counter entropy. Moreover, it’s like Gross Domestic Product (GDP) per capita. Each person has to have a GDP above 0 meaning that they are at least maintaining their current life style disregarding inflation. A person will continue to spend money on their needs. If they do not have an income or a means to sustain their life, entropy will certainly continue to diminish any assets. The higher the GDP, the better off the person is. GDP takes account for what the person has created as a finish product minus the losses of entropy. In the case of this article, finished products is work one has done that has value to him or her and others in some way either through money, love, material goods, etc.

People can lose their jobs — — or a relationship could end — — or someone could get really sick. It’s not always negligence as life has its unfair grey areas. But Entropy is consistent and people must be aware that even if things in life seem good, entropy is always at work [sarcasm]. If people are proactive, they can minimize the impact of Entropy. Entropy is the loss of energy to a subject or matter. I utilized natural decay to reflect how entropy impacts human life. Entropy as a topic can be complex and there is a whole field in thermodynamics where entropy can take different mathematical forms. For the purpose of scalability and given the complexity of the human mind, I decided to describe entropy with one function with changing parameter values as shown in the graph below.

Number of Life Domains People manage

As you may expect people can manage n number of domains in their life. Literally there could be a million domains for any given person, but it doesn’t seem practical. According to Inc.com there are seven (7) high level domains by which successful people measure themselves: Health, Family, Social, Financial, Business, Civic and Spiritual [1]. It’s safe to say most people care about these domains in some way. Even the order shows some level of Maslow’s hierarchy of need. But people don’t always think at this high level. Thinking at the macro level is indeed intentional and sometimes taxing on the mind; thus, most people rather use the bottom up approach rather than the top down.

Bottom up : System 1 :: Top Down : System 2

There are two approaches: bottom up or top down. It depends on the person’s cognitive state at any point in time. From my experience in machine learning top down is more of a system 2 slow and methodical process. Whereas, bottom up is a system 1 biased approach that could be tunneled vision. Data Iku a company who focuses on data science puts it as, “The bottom-up approach to data science tends to be unstructured and exploratory. It lets the data lead to a result, while the top-down method defines a problem to be solved and constructs an experiment to solve it.” [2] Both sides are valid strategies and supposedly makes up how the brain questions itself. System 1 seems creative yet extremely biased and system 2 seems rigid yet credible using its entire knowledge base. It is the cross road between these two where good decisions are formed. But there is also a paradigm shift in bottom up => top down. The bottom up is the way the mind has a sea of knowledge and it is trying to organize this unstructured data between unrelated ideas in an unconscious way. Whereas, the top down is thinking and forming conclusions in a conscious way.

Formulating the Outcome Equation

We have a number of domains n to account in our lives. I will list a few.:

Paying the bills
Car is functional
Exercising to keep healthy
Eating well. Choosing the foods we eat
Relationship with significant other
Maintaining a good status at work
…

There are thousands of circumstances that each person keeps in mind; therefore, machine learning would be a good fit to solve this. You cannot write code to help every person. The cost would be unsustainable and it would burnout the team trying to maintain that code. As a matter of fact, even one person would be difficult to maintain, because what matters now to a person will shift in a few weeks. There will be thousands of one-offs that is difficult for any company to keep track.

Thinkmoat Outcome Equation

I came up with this equation shown below that focuses on similar properties to thermodynamics considering energy and losses [5]. Also this equation is similar to neurology in that if certain connections in the brain are not used, the connections are lost. Basically “if you don’t use, you lose” principle. Ideally it’s more of a formula because much of this is dynamic and at many times it can’t be reproduced in production.

To break down the variables,

Outcome (life) — is an index: is the current state of one’s life based all domains n. This function is similar to GDP but for all whelms of life. A lot of people like to keep their index value at zero, which is just maintaining their current life style. But a lot of people also want their index to be greater than 0 meaning they are progressing in life. It is unlikely anyone wants to have an index below 0 at any point in time; unfortunately, this happens to everyone in various points in life.

n — the different domains in one’s life. For example: having a car, having children, a spouse, going to college, becoming CEO, getting into a life of crime. This is all of the responsibilities one has accrued or incurred in life based on their decisions they made. There could be millions of different domains.

t — time (pretty self explanatory)

Work (W) — Is the mental and physical work necessary to make decisions that lead to prosperous outcomes. Outcomes could be a deal, an end product, a good relationship, higher GDP, etc. But work could also be negative. If people do things that are detrimental to their livelihood, they are creating negative work. Example of this is getting into a fight, not showing up to work, arguing with everyone, eating fast food everyday, drunk every night and many more.

r — the rate of output through work that a person produces or receives as a gift from others. The rate r changes all of the time. The rate of return is the gains a person receives for some level of effort either in time or product. If this value is negative, it’s due to a person taking an action that doesn’t bode well.

Entropy (E) — The constant decay of a particular circumstance in a domain. Each circumstance is different; thus, each entropy coefficient will have different magnitudes. For example if someone has a pain in their fingernail and another person has stage 5 terminal cancer, of course the entropy coefficient for the person who has cancer is orders higher than the person who has pain in their fingernail.

k — is the entropy decay rate for a particular circumstance (domain). If a person has a high position like a CEO, their k-value is high. If a person who lives a life of crime, their k-value is high in getting caught. A person who is dealing with cancer, their k-value is high because it’s harder to maintain state in these circumstances.

Note the outcome takes into account the magnitude of a person’s situation and understands its context in relation to other circumstances using numerical numbers between -1 and 1 and applying a soft max. How does this work? If a person finds out they have cancer, the new circumstance will be added to the vectors W(n, t) and E(n, t). The cancer of course will have a much larger Entropy coefficient E (Attention or importance) and the decay rate k (rate of decline) is significantly higher than the rest of the circumstances. Below is simple double integral from 1 to n and from time 0 to t for both work and entropy, respectively. Again, n is the different domains or circumstances.

The area under the curve computes the entire index of positive work a person does minus entropy. But note that not all work is positive. If a person is living a life of crime, they are creating negative work on top of entropy. Some might say in the mind of the person who is committing the crime think they are performing positive work. And it might be true for that individual. But certainly the entropy would make up for the difference given the outcome index will still be negative.

Adversely if a person became a CEO of a fortune 500 company and was paid $10 million a year along with 1.4% share value in the company after being vested for five years, the positive work side of this function would be significantly higher than the entropy. Assuming the CEO did not match their life style with his or her new pay or work their body to the ground.

Some may ask well how do you determine the value of W or r and E or k? These values seem ambiguous and don’t really have context to circumstances. My answer is — these values are calculated through machine learning and then use a soft max to relate the values to each other. For instance, if we did research and surveyed thousands of people and they indicated how they rate an outcome from 1 to 10 and we also request personal information, we can run this model. When using machine learning, I look at what was the outcome and place attention on the circumstances (domains) that caused that outcome. What this will achieve is the growth and decay rates r and k after training. Afterwards, we will use another machine learning algorithm that implements gradient descent to continuously adjust the attention coefficients W(n) and E(n) vectors to determine the optimal outcome for each person each day. What this will do is once the optimal outcome is determined, we go back to the W and E vectors and see where to the person can spend their attention based on the coefficients (weights).

Computation

If you used a standard deep learning model, it would work well. But you would have to know all circumstances that could exist before training. That is virtually impossible. That is the same issue the general intelligence community is going through. What I propose is using a Bayesian Multinomial approach then if you need to adjust the circumstances, then you adjust the corpus and then invoke the training again. It may not be as accurate as a deep learning neural network but this function provides extraordinary time reduction in corpus changing and training. In addition, when dealing with optimization, accuracy is not exactly priority. The goal is to discretely determine where to spend one’s attention at a higher level of cognition.

In production, after each prediction, I would keep the results as historical data so that we can recreate the models and pick up where it left off. The only item that we need feedback is the outcome for each circumstance. Of course this type of computation is expensive and possibly take a lot of time to ramp up.

Opposition

Building tools like this has extraordinary benefits, but there is still the issue of bad actors. Some people will try to exploit people or use a tool like this to figure out plans to do things that are harmful to others. Much like other systems such as the YouTube algorithm or other platforms, this idea needs to have safe guards; an A.I. that tries to determine if a user is acting maliciously.

Conclusion

The Cofounder/Thinkmoat Outcome Model helps A.I. optimize what’s important for people and not what’s important for the A.I. itself. What’s magical about this formula is that this shows there is no one-way to live a life. A.I. meets each individual person where they are and tries to figure out how to optimize the person’s life circumstances. This is a starting point where A.I. is not giving canned responses — one size fits all scenarios. But it is delivering unique day by day plans that can help people.

References

[1] Inc.com. “7 Secretes of Successful People to Living a Balanced Life”. https://www.inc.com/jim-schleckser/seven-secrets-of-successful-people-to-living-a-balanced-life.html

[2] Data Iku https://blog.dataiku.com/top-down-vs.-bottom-up-approaches-to-data-science

[3] Wolchover, Natalie. QuantaMagazine. “Artificial Intelligence Will Do What We Ask. That’s a Problem.” https://www.quantamagazine.org/artificial-intelligence-will-do-what-we-ask-thats-a-problem-20200130/

[4] Lumen. “Exponential Growth and Decay”. https://courses.lumenlearning.com/waymakercollegealgebra/chapter/exponential-growth-and-decay/

[5] Frisby, Frank. Thinkmoat. Mathematical Equations in Machine Learning. 2019.

YouTube: https://www.youtube.com/channel/UCpY0tFtMXIM1qZGULpuEy4Q

We do not provide our results. This is just an article on what we reviewed and some of our research. Most of this information is theoretical. We are only proposing ideas as described above. Original document — Credit to Frank Frisby.

Original source of the this article: https://www.linkedin.com/pulse/optimizing-outcomes-peoples-lives-mathematical-approach-frank-frisby/

Frank Frisby / Thinkmoat: https://www.franktfrisby.com/

Twitter: https://twitter.com/Franktfrisby

YouTube: https://www.youtube.com/channel/UCpY0tFtMXIM1qZGULpuEy4Q

LinkedIn: https://www.linkedin.com/in/franktfrisby/