Contribution
It named the exploration–exploitation trade-off as a general problem of adaptive organizations and showed, in a simple model, that the speed at which an organization socializes its members is itself a choice about exploration, with costs and benefits distributed unevenly among members.
Key insight
An organization’s store of knowledge can learn only from the people who disagree with it. Teaching them its beliefs quickly makes each of them more accurate sooner and makes the organization stop learning.
Assumptions
- Reality has m binary dimensions (m = 30); n = 50 members and a code hold beliefs of 1, 0 or −1 on each.
- Members learn only from the code (probability p1 per differing dimension); the code learns from the majority of the members who know more than it does (rate governed by p2). Members do not learn from each other directly, and nobody observes reality.
- Turnover (p3) replaces members with recruits whose beliefs are random, not drawn toward the code. Turbulence flips dimensions of reality with a small probability per period.
- 80 runs per setting; n is not varied. March says the qualitative results are insensitive to m and n.
Key results
- Closed system (p. 75). Members and code converge to a stable equilibrium of shared, not necessarily accurate, beliefs. Equilibrium knowledge is highest when the code learns fast from members socialized slowly; with fast socialization, slower code learning does better.
- Heterogeneity (pp. 76–77). For any average p1, a mix of fast (0.9) and slow (0.1) learners reaches higher equilibrium knowledge than a uniform population. Gains come from slow learners and accrue to fast ones, so no individual has an incentive to learn slowly.
- Turnover (pp. 79–80). With fast socialization, moderate turnover improves the code; with slow socialization, fast turnover leaves too little exploitation.
- Turbulence (pp. 80–81). Without turnover, code knowledge rises and then falls to chance, a random walk. Moderate turnover (p3 = 0.1 with turbulence at 0.02) sustains it. Recruiting people closer to the code would weaken this. Tenure then advantages individuals, which March expects to produce pressure to secure tenure for oneself and restrict it for others.
- Competition for primacy (pp. 81–84). As competitors increase, variance in performance matters more and, in the limit, “the mean becomes irrelevant”. Knowledge that raises mean and reliability does not guarantee advantage.
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | Adaptive processes improve exploitation faster than exploration, and can become self-destructive | moderate | argument from the timing and certainty of returns; prior literature on competency traps |
| C2 | Slow socialization raises equilibrium organizational knowledge, especially when the code learns fast | moderate | simulation, Figure 1 |
| C3 | A mix of fast and slow learners beats a uniform population with the same average rate | moderate | simulation, Figures 2–3 |
| C4 | Moderate turnover improves organizational knowledge, and under turbulence prevents its decay to chance | moderate | simulation, Figures 4–5; depends on recruitment not being drawn toward the code |
| C5 | Where only relative position counts, variance can matter more than mean | moderate | model of independent draws; acknowledged incomplete |
Method
Monte Carlo simulation of a stylized model, 80 runs per parameter setting, with m = 30 and n = 50. Outcomes are the proportion of reality correctly represented in the code and, on average, in members’ beliefs, at equilibrium or at period 20. The competition section uses a separate analytic model of performance as independent draws from distributions.
Concepts
- exploration — “experimentation with new alternatives”; returns “uncertain, distant, and often negative” (p. 85).
- exploitation — refinement and extension of existing competences; returns “positive, proximate, and predictable” (p. 85).
- organizational code — the languages, beliefs and practices of the organization; its knowledge level is the proportion of reality it represents correctly.
- socialization rate (p1) — how fast members adopt the code.
Connections
It cites earlier work by Herriott, Levinthal and March on slow learning, and David on path dependence. Its model resembles the epistemic-network models behind THEORY-148: in both, diversity in what members believe or pursue must last long enough to be learned from. The mechanism differs. Here diversity is lost to a central code; there, to communication among peers.
Bearing on the record
Source of THEORY-142.
For the owner’s essay: §5.4 (protective lock-in) gets a mechanism by which an organization’s beliefs stay fixed while the world changes, “regardless of changes in reality”, without anyone deciding it. §6.1 gets a case of stability that is not viability: the converged equilibrium is stable and degrades. §8.2 gets an organizational analogue of Longino’s and Zollman’s point that a community’s knowledge depends on keeping dissent alive. The model has no exit, voice or governance: turnover is an exogenous rate, not members choosing to leave, so it says nothing about how members respond to decline.
Limitations
Results come from one stylized model with fixed m and n. Members never observe reality, and never learn from each other. Turnover’s benefit depends on recruitment rules March sets aside. The competition analysis assumes independent draws and is called “incomplete” because mean and variance can be chosen strategically. No optimal balance is derived.
Open questions
Whether the results hold when members learn from each other as well as from the code, and what organizational structures, other than turnover, keep deviation alive long enough for the code to learn from it.