Observations of Society

Main Menu           Previous Topic                                                           Next Topic

The Effect of Learning Rates

We now turn to the question of how we can optimize the knowledge in an organization. Of course, the answer will depend on the specific sense in which we define an optimum. Different strategies will create different distributions of knowledge across time. The equilibrium between the agents and the code will generally appear faster with higher learning rates. What about the amount of knowledge in society when the equilibrium occurs? In the graph below, we plot equilibrium levels of knowledge found in society under a variety of socialization rates and code learning rates. Each data point is based on 1000 trial societies. The population size is held at 50; the number of dimensions of reality is fixed at 30.

Clearly, faster rates of learning are not always optimal. When the code only adapts to agent beliefs at a slow rate, we do see that faster socialization rates lead to higher equilibrium knowledge. However, for moderate levels of code learning, knowledge is optimized by slow agent learning. When socialization is slow, equilibrium knowledge is increased by increasing the code learning rate. However, when socialization is rapid, fast learning by the code improves equilibrium knowledge. The highest equilibrium knowledge is achieved by slow learning by individuals from a code that learns quickly.

Why would it be advantageous for agents to learn slowly? According to March, the immediate gains to individuals from adapting quickly to the code are offset by second-order losses that occur because the code can only learn from agents that differ from it. The code benefits from diversity.

Thus, in this situation, there is a difference between a social equilibrium, in which agents wish to learn as quickly as possible from the more intelligent code, and the social optimum, which favors the diversity that is preserved longer by slow socialization. Fast learning by agents corresponds to exploitation, yielding quick immediate gains for socialized individuals. Slow learning corresponds to exploration, maintaining a diversity of beliefs. The equilibrium state in the latter case is more accurate, but is achieved more slowly.

As demonstrated in this scenario, an organization may implicitly limit the amount of exploration it conducts by seeking out fast learners, or providing incentives to agents to socialize into the code.

                   Previous Slide                                                           Next Slide