On consensus and compromise decisions
Positions pro and aganst consensus. Analogy between consensus decisions and ensemble learning.
In this post, I aggregate what different people wrote about consensus and compromise decisions.
In his courses, Earl Beede calls consensus one of the good decision-making processes, alongside delegating decisions to implementers and “decider after a discussion”. Steve McConnell (the author of Code Complete and Software Estimation who works with Earl Beede at Construx Software) also suggests to try to find a “golden mean” third alternative in his paper “Managing Technical Debt” (see summary):
Teams often turn the technical debt decision into a simplistic “two option” decision—good path vs. quick and dirty path. Pushing through to a third option is important because often the best path is the one that is fairly quick, albeit not as quick as the pure quick and dirty path, and whose adverse effects are better contained than those of the pure quick and dirty path.
The idea of a golden mean dates back at least to Aristotle.
Think through what is right: the solution that will fully satisfy the specifications before attention is given to compromises, adaptations, and concessions needed to make the decision acceptable. You will always compromise in the end.
In The 7 Habits of Highly Effective People, Stephen Covey writes that despite being worse than truly creative, synergistic cooperation (as David Deutsch suggests—see below), consensus is better than any one's position alone:
In interdependent environment, respectful communication leads to compromise positions, or pleasantly enforced consensus. It's honest cooperation, but not creation. Transaction, not transformation. 1 + 1 = 1.5.
David Graeber makes an interesting note in Debt:
Buddha admired democratic organisation of North India city states and adopted it for organisation of his followers. Buddhist monasteries continue to operate by consensus to this day.
Analogy between consensus decisions and ensemble learning
In machine learning, some ensemble learning methods are equivalent to consensus. Ensemble methods typically make better predictions and are better regularised than individual models used in the ensemble.
Richard Hamming in The Art of Doing Science and Engineering:
Committee decisions diffuse responsibility. Most of the time they are compromises that have none of the virtues of any path.
David Deutsch agrees with Hamming in The Beginning of Infinity:
Good explanations are hard to vary, so a mix of two is usually worse than both. Combining two explanations requires an additional act of creativity.
Large groups of people look for consensus, not truth.
Carmen Medina in “Rebels at Work”:
Consensus is a way to avoid making a decision, when you talk to the lowest common denominator.
Making a decision is an act of creativity
Complex problems require creative solutions because consensus is often impossible in a complex situation.
David Deutsch again:
Decision making is really problem solving. Without creative process, nothing is ever solved and there is nothing to choose from. At the heart of decision making process is creating new options and abandoning or modifying new ones.
To choose an option, rationally, is to choose an associated explanation. Therefore, rational decision making consists not of weighing evidence, but of explaining it in the course of explaining the world. One judges arguments as explanations, not justifications.
Typically, one struggles to create even one good explanation, and once created it, gladly discards the rest.
According to Deutsch, complex decisions have a long "perspiration" phase when vague details are specified through non-creative work. But it begins with answering a big question creatively, such as "How client's priorities and constraints can best be met by the new design?"
Deutsch argues for all-or-nothing political voting system because it makes the politicians to find the best explanations and convince most people of them, and when they win, they are solely responsible for the result. Proportional parliament creates compromise policy which is bad because nobody learns anything if it fails.