Starcraft, Jaynes, and Bayesian UQ

Interesting content covered on Nuite Blanche of a recent Paris Machine Learning meetup. The work on applying Bayesian Programming and Learning for Multi-Player Video Games was really neat. It's about developing a bot for playing Starcraft. Some additional related presentations:

• A Bayesian Model for Plan Recognition in RTS Games applied to StarCraft
  
• A Bayesian Model for RTS Units Control applied to StarCraft
  

This problem is real enough to be interesting, and simple enough to tackle. I really liked how this work dealt with the separation of strategic/tactical level decision making.

Efficiently Directing the Work

The good Colonel knew it in 1918,

no man can efficiently direct work about which he knows nothing
--Col Thurman H. Bane

and we can rediscover it nearly a century later,

Item 19 of the checklist stresses the importance of placing experienced, domain-knowledgeable managers in key program positions. The committee has observed that many of the truly extraordinary development programs of the past, such as Apollo, the Manhattan Project, the early imaging satellite programs, the U-2, the fleet ballistic missile system, and nuclear submarines, were managed by relatively small (and often immature) agencies with few established processes and controls. In that environment, dedicated managers driven by urgent missions accomplished feats that often seem incredible today.

The committee believes that the accumulation of processes and controls over the years—well meant, of course—has stifled domain-based judgment that is necessary for timely success. Formal SE processes should be tailored to the application. But they cannot replace domain expertise. In connection with item 19, the committee recommends that the Air Force place great emphasis on putting seasoned, domain-knowledgeable personnel in key positions—particularly the program manager, the chief system engineer, and the person in charge of “requirements”—and then empower them to tailor standardized processes and procedures as they feel is necessary.

[...]

While the systems engineering process is, broadly, reusable, it depends on having domain experts who are aware of what has gone wrong (and right) in the past recognize the potential to repeat the successes under new circumstances and avoid repeating the errors.

Pre-Milestone A and Early-Phase Systems Engineering: A Retrospective Review and Benefits for Future Air Force Acquisition

When I read that last part it reminded me of something Herbert Mason said at a talk he gave recently at the NMUSAF: "History makes you smart, heritage makes you proud."

Only Opportunity

Happened upon this

The constraints imposed by the planetary ecosystem require continuous adjustment and permanent adaptation. Predictive skills are of secondary importance.
Hendrik Tennekes

 and thought of this

There is no security on this earth; there is only opportunity.
General Douglas MacArthur

Tennekes concludes with

From my background in turbulence I look forward with grim anticipation to the day that climate models will run with a horizontal resolution of less than a kilometer. The horrible predictability problems of turbulent flows then will descend on climate science with a vengeance.

I have his book; it is rather good.

Zen Uncertainty

Zen Uncertainty: Attempts to understand uncertainty are mere illusions; there is only suffering.
-- WARNING: Physics Envy May Be Hazardous To Your Wealth!

Should we give up? No, there's plenty we can do to make the suffering more bearable. Lo and Mueller give an uncertainty taxonomy of five levels in their 'Physics Envy' paper:

1. Complete Certainty: the idealized deterministic world
2. Risk without Uncertainty: an honest casino
3. Fully Reducible Uncertainty: the odds in the honest casino are not posted, we have to learn them from limited experience
4. Partially Reducible Uncertainty: we're not quite sure which game at the casino we're playing so we have to learn that as well as the odds based on limited experience
5. Irreducible Uncertainty: we're not even sure if we're in the casino, we might be outside splashing around in the fountain...

At the bottom of the decent we find level infinity, Zen Uncertainty.

Section 2 of the paper provides a nice historical overview of the early work of Paul A. Samuelson, who single-handedly brought statistical mechanics to the economists, and they have never been the same since. Samuelson acknowledged the deep connection between his work and physics:

Perhaps most relevant of all for the genesis of Foundations, Edwin Bidwell Wil- son (1879–1964) was at Harvard. Wilson was the great Willard Gibbs’s last (and, essentially only) protege at Yale. He was a mathematician, a mathematical physicist, a mathematical statistician, a mathematical economist, a polymath who had done first-class work in many fields of the natural and social sciences. I was perhaps his only disciple . . . I was vaccinated early to understand that economics and physics could share the same formal mathematical theorems (Euler’s theorem on homogeneous functions, Weierstrass’s theorems on constrained maxima, Jacobi determinant identities underlying Le Chatelier reactions, etc.), while still not resting on the same empirical foundations and certainties.

Related to this theme, there's an interesting recent article over on Mobjectivist site about using ideas from physics to model income distributions.

Lo and Mueller propose to operationalize their uncertainty taxonomy with a 2-D checklist (table). The levels provide the columns across the top, and there is a row for each business component of the activity being evaluated, here's their description:

The idea of an uncertainty checklist is straightforward: it is organized as a table whose columns correspond to the five levels of uncertainty of Section 3, and whose rows correspond to all the business components of the activity under consideration. Each entry consists of all aspects of that business component falling into the particular level of uncertainty, and ideally, the individuals and policies responsible for addressing their proper execution and potential failings.

This seems like an idea that could be adapted and combined with best practices for model validation (and checklist sorts of approaches) in helping to define what sorts of uncertainties we are operating under when we make decisions using science-based decision support products.

Their final paragraph echos Lindzen's sentiments about climate science:

While physicists have historically been inspired by mathematical elegance and driven by pure logic, they also rely on the ongoing dialogue between theoretical ideals and experimental evidence. This rational, incremental, and sometimes painstaking debate between idealized quantitative models and harsh empirical realities has led to many breakthroughs in physics, and provides a clear guide for the role and limitations of quantitative methods in financial markets, and the future of finance.
-- WARNING: Physics Envy May Be Hazardous To Your Wealth!

Joint Targeting Zen

Sometimes the most important part of the targeting cycle is deciding what targets not to engage to achieve the effects we want.

"It's possible (a strike) could be used to play to nationalist tendencies," Petraeus, head of the U.S. Central Command region, which includes Iran, said in an interview this week. "There is certainly a history, in other countries, of fairly autocratic regimes almost creating incidents that inflame nationalist sentiment. So that could be among the many different, second, third, or even fourth order effects (of a strike)." --Patraeus Says Strike on Iran Could Provoke Nationalism

The response of the British populace to The Blitz provides a good historical example of the kind of thing Gen Patraeus is describing.

Thirty spokes
Round one hub.
Employ the nothing inside
And you can use a cart.
Knead the clay to make a pot.
Employ the nothing inside
And you can use a pot.
Cut out doors and windows.
Employ the nothing inside
And you can use a room.
What is achieved is something,
By employing nothing it can be used.
--Tao Te Ching, 11

Fog and Friction

From the short story by C.D.B. Bryan.

From these things the army was taught the nature of the place and how the slope of the hill and the necessity to the time demanded more than one plan and order for the art of war. Different legions, some in one part, others in another, fought the enemy. And the view was obstructed by very thick hedges. Sure support could not be placed, nor could it be seen what work would be necessary in which part, nor could all the commands be administered by one man. Therefore, against so much unfairness of things, various consequences ensued.