More reasons defense acquisition needs to consider randomness

Discussing Tim Harford’s book Messy, he tells a story of a Jazz musician Keith Jarrett who, upon a major performance was given a poor piano, was able to improve on it and made one of the great jazz albums of all time. Here’s Tim:

So, a few things to say about that. One is: Obstacles sometimes help us solve problems. Two, is, we don’t expect to see that in advance… Various pieces of research from cognitive psychology show that when we are kind of knocked sideways or distracted or given slightly more tricky tasks, unnecessarily tricky tasks, we often produce a more creative response.

 

We can see this through the lens of computer algorithms. A computer algorithm that has to try to find an approximate solution to a very complex problem. There are a whole family of these algorithms, but they all have one thing in common, which is they all tend to use some randomness, especially early on. And the reason they use randomness is because otherwise they get stuck at a local optimum. They go down some blind alley; they find a spot that’s pretty good but not all that good. And they can’t get out. And this search program, the computers using, can’t get out of this blind alley. But if you throw randomness in and the computer suddenly makes a big move and tries something totally different, then it gets unstuck.

 

So, both psychology and computer science suggest that we should not be entirely unhappy when something unexpected happens, when we are forced to work with a difficult person or with a substandard tool. It’s not necessarily a problem.

I think that makes a lot of sense. In defense acquisition, every program is expected to be perfectly correct from the outset — no errors or pivots in cost-schedule-technical. Yet that means defense officials are institutionally incentivized to prefer the least risky programs, as opposed to the programs with the greatest possible impact.

Then, the continual desire to remove duplication and overlap from programs creates a blind alley for DoD. Inward looking systems tend to become confused and disordered.

A couple major things DoD needs to change to inject some of that experimental randomness:

• Higher threshold for New Starts. Imagine if a computer system had to ask the bureaucrats and Congress every time it wanted to start its conjugate gradient method at a different point to find the global maximum? “Is this coordinate a likely place to find the global maximum?” How would they even know? That’s why systems use randomness.
  • Raise threshold for congressional notification from $10M over life of the project to $10M for the use of a year’s appropriation. Consider fewer restrictions over time.
• No Full Funding Requirements. The idea that you’ll know precisely the lifecycle cost of a complex RDT&E and Procurement program again presumes the end-state can be known — which means randomness is the devil. But in the real world of uncertainty and accelerating technology, there has to be some indeterminacy of lifecycle costs. The 5-year future should be overpacked with programs, because not all of them should be expected to work out perfectly, or perhaps they merge and cooperate in unexpected ways.