Modeling Everyday Learning: From Reinforcement to Social Learning

Yesterday I realized that many of the things I study are fundamentally about learning, not just formal education but also everyday decisions like commuting, ordering food, or playing games.

Learning is essentially a process of continuously accumulating experience and adjusting strategies to better adapt to the environment. It involves not only receiving external information but also filtering, processing, and converting it into actionable behavior, constantly refining decisions through feedback.

The process includes two key aspects:

Exploration : trying new things and gathering more information, which involves risk and possible failure.

Exploitation : using existing experience to choose the currently optimal option, emphasizing rational decisions based on past rewards.

Balancing exploration and exploitation makes learning a dynamic adjustment process that seeks optimal solutions in new contexts.

When I thought further, I saw that these concepts can be expressed with mathematical models, the first that came to mind being reinforcement learning, though learning also occurs socially.

Reinforcement Learning Model

When choosing where to eat, I tend to repeat restaurants that previously gave good food and hygiene, which is a reinforcement learning process.

We can describe this with a simple mathematical model. Suppose there are three candidate restaurants with initial equal preferences (weights). The probability of selecting a restaurant is proportional to its weight. After receiving a reward, the weight is updated using:

weight_new = weight_old + learning_rate * (reward - expected_level)

where the learning rate controls adjustment speed and the expected level is the average reward.

Repeated trials reinforce preference for high‑reward restaurants, eventually converging to the optimal choice.

This model captures individual decision‑making but cannot account for unseen options that might offer higher rewards, leading to a bias toward known choices.

Social Learning Model

Consider choosing between two driving routes. Initially you may pick randomly, but observing many others taking one route and arriving earlier may cause you to switch—this is social learning.

We can describe it with a replicator dynamics model:

Assume two routes with different rewards.

The probability of choosing a route changes proportionally to its relative reward compared to the average.

Over time, the population converges on the higher‑reward route.

Unlike reinforcement learning, social learning incorporates others' behavior, making it suitable for decisions in social contexts.

In everyday life, we often rely on reviews or recommendations, which is another form of social learning.

Social learning explains collective behavior and herd effects, such as stock market convergence, but it can also lead to suboptimal “herding” where the majority adopts a non‑optimal choice.

Insights for Learning

Reinforcement learning emphasizes individual experience and suits independent decisions, while social learning highlights interaction and imitation, better explaining group behavior but risking herd effects.

Combining both models can provide a more comprehensive explanation of real‑world learning, useful in investment decisions, product selection, traffic management, and other domains.

Learning is the core ability for humans and other organisms to adapt to environments and achieve goals. Mathematical modeling lets us quantify and analyze seemingly complex learning behaviors.

In the future, learning will involve not only individual‑environment interaction but also mutual learning among people, machines, and machines themselves, becoming essential for thriving in the intelligent era.

A person’s most important habit to develop is the desire to keep learning.

Reference: Scott Page, Model Thinking: How to Make Sense of the World (2023).