Systems Whose Parts Learn and Adapt

The idea

In an ordinary machine, the parts follow fixed rules. In an adaptive system, the parts learn. Each one watches what the others do and changes its behavior, so the rules of the whole system drift as its history unfolds. Patterns appear that nobody planned, a theme I explore in Studying the Parts or the Whole.

Why it attracts me

It explains why so many good plans decay. An optimal answer (Finding the Best Plan Within Limits) is optimal for the system as it was when I measured it. Once people respond, I am solving a different problem. That makes watching and adjusting as important as the first calculation.

Where it connects

This is where systems meet games. When the parts are people with their own goals, each choice depends on what others choose, and habits and norms can spread or fade as conditions change (How Behaviors Spread or Fade in Groups). The influence runs both ways: people's choices reshape the system, and the system reshapes their choices. It also sharpens how I think about ripple effects (Feedback Loops and Ripple Effects).

An example

Imagine a support team that is rewarded for closing tickets quickly. At first, response times improve. Then people learn to split big jobs into many small tickets, or close tickets early and reopen them later. The measure still looks good, but the system has adapted around it. Nothing broke; the participants simply learned. A better design would expect that learning from the start, and would keep checking whether the measure still means what it meant on day one.

Questions I'm still exploring

• How can I tell whether a system has adapted to my change or simply drifted on its own?
• When is a fixed best plan good enough, even though people will adapt to it?
• Can a simple computer model of learning players say anything trustworthy about real people?

Sources and further reading

• John H. Holland, Hidden Order: How Adaptation Builds Complexity (1995)