Feb. 04, 2025
Why Is Toyota Conducting Brain Research: volume 4Strategies for Promoting Collective Well-being through Our Actions
The Frontier Research Center of Toyota Motor Corporation (hereafter, Toyota) is exploring the possibilities created by integrating brain science and engineering: the aim is to generate innovation for the future society. As a part of that process, in 2007, Toyota established an organization for comprehensive collaboration, the RIKEN CBS-TOYOTA Collaboration Center (BTCC), with the RIKEN Center for Brain Science (RIKEN CBS) in Wako, Saitama Prefecture. In the fifth phase of BTCC, which began in the 2023 fiscal year, we aim to construct a framework and design theory for the synergistic development of individual and collective well-being*1. In this interview, we spoke with Dr. Wataru Toyokawa, who leads the Computational Group Dynamics Collaboration Unit at BTCC, about the latest research.
- Many people experience that imitating others in sports or fashion can enhance individual well-being. Dr. Toyokawa, you study such imitation behaviors; how important are they scientifically?
- Dr. Toyokawa
- Looking at living organisms in general, not all species engage in imitation, but many animals do. For example, fish, birds, and primates like chimpanzees and Japanese macaques are known to imitate the behaviors of other individuals. Such imitation forms the basis of the collective behavior patterns of those animals.
- I see, imitation significantly influences group behavior.
- Dr. Toyokawa
- Exactly. When someone starts showing one behavior that works very well, it gets imitated by others and spreads across the population. Human imitation behavior is particularly interesting because we are known to mimic others very faithfully. Imitation plays a crucial role in the formation of human culture, including language, the use of tools, and establishing various social systems.
- Imitation is an essential element not only for our daily well-being but also for the development of culture and society. By the way, although we all have different values and preferences, does it always make sense for us to imitate each other?
- Dr. Toyokawa
- Good question! In groups composed of individuals with different values and preferences, it's not obvious whether individual imitation behaviors contribute to the group's benefits. To clarify this, we conducted an experiment using the game shown in Figure 1.
-
- Figure 1: Social Learning Experiment Task Using a Planet Exploration Game
The left side of the figure shows the amount of salts found in searched locations. The reward map on the right indicates that areas with darker yellow contain more salts, while areas with darker blue contain very few salts.
- Dr. Toyokawa
- In this game, four players act as an exploration crew on a planet searching for salts, trying to maximize the total amount of salts each player collects. The target salts differ for each player, and at the outset, it is unknown for players where and how many are buried on the map. Players can check the history of how many salts were found in the locations they searched. Additionally, all players were given the hint that "places where many salts can be found are somewhat similar, regardless of the type of salt."
- So, when other players find a lot of salts in a location, players think they might also find a lot there.
- Dr. Toyokawa
- That's right. However, not all salts are distributed in the same way. Look at the reward map of Crew 1 shown on the right in Figure 1. It shows that areas with darker yellow contain more salts, while darker blue areas contain very few. Comparing Crew 1's reward map with Crew 2 to Crew 4's maps, you can see they are somewhat similar but not identical. This creates a situation where each player has different values, and we confirmed how players learn from their own choices and outcomes as well as from the choices and outcomes of others (social learning) through simulations and human experiments.
- Through simulations, we can investigate how individual imitation behaviors occur. Could you elaborate on that?
- Dr. Toyokawa
- Imagine a virtual space created by a computer filled with many agents playing this game. For example, some agents do not refer to other players at all, while others imitate based on certain learning models. We allow these agents to compete to achieve high game scores. Agents with higher scores are more likely to reproduce, passing on their behavioral patterns to their offspring.
After many generations, the logic of natural selection reveals which strategies persist. This allows us to identify which behavioral rules (that is, learning models) perform well in the given environments. We call this evolutionary simulation.
In this evolutionary simulation, we compared four learning models. The first is individual learning that does not refer to others at all, the second and third are learning models commonly used in traditional social learning models, and the fourth is the social generalization learning model we proposed.
- What is the social generalization learning model?
- Dr. Toyokawa
- Figure 2 represents a situation where Player A found a quantity of 82 salts in a location. If Player A's reward map is similar to mine, similar amounts of salts might be found in the same location or nearby.
Our proposed social generalization learning model considers information obtained from others as "less reliable experiences" and generalizes it, applying concepts from Gaussian process learning models studied in cognitive science and machine learning.
-
- Figure 2: Generalization Using Gaussian Process Learning
- What did the simulations reveal?
- Dr. Toyokawa
- Look at the left graph in Figure 3. The horizontal axis represents social correlation coefficients. When this value is 0, the reward maps of the players are not similar at all; when it is 1, the reward maps are identical. The vertical axis shows the frequency of evolutionarily successful strategies, indicating how prosperous a group with a specific learning model was.
-
- Figure 3: Analysis of Evolutionary Simulation and Experimental Data
In the left graph, DB (yellow line) and VS (green line) appear to be completely identical, making DB seem invisible.
- Dr. Toyokawa
- From this result, we can derive two major conclusions. First, when the social correlation coefficient is small (around 0 or 0.1), groups that engage in individual learning thrive the most. This means social information is not very reliable. Regardless of what others do, it does not serve as a good reference, so in such cases, it's best to work alone. This result may seem very intuitive and obvious.
The second conclusion is that when the social correlation coefficient is 0.2 or 0.3 or higher, the proposed learning model―the social generalization learning model―functions well. In these ranges, although the reward maps do not appear to be similar at first glance, social learning can enhance decision-making performance mutually, leading to improved overall group performance. This is such a surprising result and I was very excited when I first saw it.
- That's fascinating! But how could social learning still work well and contribute to group performance even though individuals do not 100% share their goals? Did you find anything in human experiments?
- Dr. Toyokawa
- In the experiments, we conducted the same game in two patterns: one where a single person played and another where four people participated simultaneously under conditions with a social correlation coefficient of about 0.6. We examined which learning model best explained people's learning strategies in each case.
The right graph in Figure 3 shows these results. The horizontal axis represents candidate learning models, and the vertical axis indicates the likelihood of each learning model being adopted by people. When playing alone, it was evident that individuals were most likely using the individual learning model. This is a very natural result.
- How about when multiple people participate simultaneously?
- Dr. Toyokawa
- In situations where multiple people participated simultaneously, meaning social learning was possible, our proposed social generalization learning model fitted the data best. However, caution is needed in interpretation. This does not mean we fully understand the human learning model scientifically; rather, it indicates that among the four learning models compared, it was the one most similar to human behavior.
- I see, scientifically, this is a step forward. Can you share your outlook on future research?
- Dr. Toyokawa
- Real societies are much more complex than what we studied, and I want to approach that complexity. For example, in our study, we maintained the social correlation coefficient, but in real life, there are people who are very similar to you and others who are completely different. It raises interesting questions about whether it's better to prioritize information from people similar to you, or if there is value in referencing information from those who are quite different.
Additionally, our study assumed that multiple players could see each other, but human societies are larger, and we cannot see everyone else. We can see friends or those we are connected to in some way, but we cannot directly see those who are not connected. I believe the structure of networks significantly impacts how information spreads*2.
- Thank you for your valuable insights. I look forward to seeing how research on imitation behavior contributes to people's well-being in the future.
Contact Information (about this article)
- Frontier Research Center
- frc_pr@mail.toyota.co.jp