Unified Artificial Intelligence

Noah Goodman and others at MIT have recently developed an A.I. based program for determining who would send e-mail to whom at a fictitious company.  However, it’s not the application itself that is of interest.  It’s the way that they carried out the computations.  The program relies on probabilistic rules that get updated over time.  It combines some of the original ideas of expert systems (using rules and implications) with the probabilistic approach that has been successful in recent years.

The research team admits that it doesn’t have a final solution on how this should be accomplished.  It currently is too computationally intensive.

Here are some random thoughts on this issue.

1. This sounds like an ideal topic for O.R. researchers to get involved in.  We also have expertise in logic and probabilistic reasoning.
2. Humans are terrible at dealing with probabilities but we are pretty good at forming categories and recognizing patterns.  Perhaps the A.I.  system could be improved by using “flawed human reasoning.”  For example, the system could be overconfident, just as humans are.  It could rely too heavily on recent information.  It could use very simple rules for updating probabilities.  It could rely heavily on other program’s expertise, just as we rely on what other people think.
3. Perhaps the best way of developing a “unified approach” is to concurrently use several approaches to arrive at conclusions, and refer to each approach as “an expert.”  Then use a unifying approach that takes the expert opinions and arrives at a group opinion.  I think that this is already done, but perhaps in slightly different ways than I am suggesting.

The Health Care Debate and Basic Research

There are two fundamental issues dealing with health care in the U.S.:  coverage and costs.  Currently, far too few people are covered by health care plans, and the costs of health care delivery is very high and increasing rapidly.

In the recent health care debate, the Democrats focused almost entirely on coverage issues, while making efforts to “do no harm” with respect to costs.   They developed an imperfect bill, with many flaws, but which is a very good step forward in dealing with the issue of coverage.  On the other hand, the Republicans were not responsible in their approach to this issue.  They appealed to people’s fears in ways unbecoming of the opposition party.  They talked about “killing grandma, the loss of freedom in America, and killing babies, all of which were based on lies.  They pretended to be the defender of Medicare, which showed amazing chutzpah on their part.  They also focused on secondary issues such as complaints about the process, lack of bipartisanship, complaints about the length of the bill, and so on.  Although they did talk about serious policy issues such as tort reform and increasing taxes, the serious issues were lost in the fog of their secondary complaints and their fear mongering.

Hopefully, we can come together and address the issue of maintaining costs in a serious and productive manner.  Everyone is affected by the costs, which are increasing much faster than inflation.    Operations Research can definitely help with this issue.  O.R. has a long history of improving the efficiency of systems, including many efforts in health care.    But I think that we can do much more, especially when we work closely with those outside of our field.

Serious efforts require serious funding.   I propose that the government spend from $5 billion to $10 billion per year in research   that is dedicated to reducing the cost of health care while improving health outcomes.  This reflects a .5% to 1% proportion of our federal government’s spending on health care.   A one-time 1% sustained reduction in health care costs would pay for this program in perpetuity.   A much larger reduction is likely.

Here are some questions that the government funding agency may wish to consider?

1. What can government do to reduce the costs of bringing an important drug to market?  If they dramatically reduce the cost of bringing a drug to market, how can they ensure that the drug companies will keep the costs of the drug reasonably low?
2. How can hospitals be run much more efficiently?
3. What less expensive alternatives are there to hospitals for people who do not require hospital care, but do require monitoring and help?  How effective are these alternatives?
4. How can electronic medical records be maintained at a low cost while ensuring the right level of privacy?
5. What medical screening tests can be developed that will dramatically reduce the costs of current testing procedures?  Here I am thinking of tests that are cheap, and have few false negatives.  People who test positive would be given the more expensive and more reliable tests.  People who test negative would need no further tests.   This would improve costs dramatically if the diseases were rare and if most test results were negative.
6. How can we reform our tort system while ensuring that patients who receive harmful and incompetent care are compensated for losses?  How can we ensure that incompetent doctors lose their license to practice?

The list can go on and on.   Health care offers so many opportunities for improved efficiencies.  I hope we will take full advantage of these opportunities.  We certainly need to.

The game theoretic advantage goes to the Republicans

Pundits have asked themselves why it is so difficult for the Democrats to get things done even with 60 votes in the Senate, whereas the Republicans seem to do just fine with 50 votes in the Senate.  I suspect that there are a number of reasons for this.  But one reason that deserves mentioning is that Democrats believe in the importance of the federal government, and Republicans believe in the importance of thwarting Democrats.

Consider the following table, which illustrates the values to Democrats and Republicans when the Democrats are in majorities in Congress.

	When Democrats control congress
	keep government going	gridlock
Republicans	1	9
Democrats	9	1

The numbers are made up.  But the key element is that Republicans would rather see nothing done at all than seriously compromise with the Democrats to pass legislation.

On the other hand, consider the reverse situation in which Republicans are in charge.

	When Republicans control Congress
	keep government going	gridlock
Republicans	7	3
Democrats	7	3

Here the Democrats really don’t want government to shut down, and they are willing to go far more than half way to meet the Republicans.  In this case, even the Republicans are willing to compromise a little.  For example, the Republicans were willing to greatly increase the National debt under Bush rather than having permanent gridlock so long as the wealthiest Americans could get a tax cut and Bush was able to start two wars; however, on other important issues such as having two ultra-conservative judges appointed to the Supreme Court, they stood by their principles and were even willing to get rid of filibusters.

I’m not sure what the moral of this story is, except perhaps that is sucks to be a Democrat.

How to make mammogram testing cheaper and more effective.

New recommendations concerning mammogram testing have upset a lot of people.   Previously, doctors recommended that women be tested yearly starting at age 40. The new recommendations say that most women should start mammogram testing at age 50, and only do it every other year.  The analysis relied on the following empirical result: one cancer death is prevented for every 1,904 women age 40 to 49 who are screened for 10 years, compared with one death for every 1,339 women age 50 to 59, and one death for every 377 women age 60 to 69.  (These recommendation make sense, as pointed out by Mike Trick.)   However, if the ratio of cancer deaths prevented per mammogram could be improved for women aged 40, then mammograms would be recommended.

What this highlights is the importance of a cheap preliminary test with reliability that is pretty good.   For example, suppose that there was a test for which no women with breast cancer had false negatives, and where 1 out of 10 women with no breast cancer had a false positive.  With such a test, approximately 90% of women would be screened out as negative, with all of them having no cancer.  The remaining 10% would test as positive, with the vast majority of them being false positives.  These women would be given mammograms.  (Women would need to be reassured that the preliminary test is not a test for cancer but a test on who should be screened using mammograms.)    Since only 1/10^th as many women would be given mammograms, this would result in cancer death prevented for every 190 women being tested, a huge improvement over the current system, with huge savings in terms of health outcomes and dollars.

Even if the test had a false negative rate of 10%, the prevention of cancer deaths would be almost as high.  There would be 1 cancer death prevented for approximately 210 mammograms.  Unfortunately, the false negative could lead to 1 cancer death for every 17,000 women not given mammograms (because of a negative test.)   At the same time, it would eliminate the health risks of mammogram testing for 17,000 women who didn’t have cancer.

The conclusion is that a cheap test with pretty good reliability (say around 90%) can dramatically improve the health outcome of mammogram testing,  while dramatically decreasing the cost of health care.   The NIH should dedicate research funding to find such tests.

The advantage of making papers harder to understand

When I was a college student, a professor of logic once asked us to write a brief paper on a specified topic, which I no longer remember.  I do remember that I didn’t understand the question that we were supposed to address and had no clue what to write.  So I carried out an experiment.  I wrote something such that each sentence appeared to make sense, but so that the paper did not make any sense when taken as a whole.  (Think “Sarah Palin’s speeches” but with crisper logic.)  My experiment was successful as far as I was concerned.  I got a “B.”  My conjecture is that it would have taken the professor too much time to see through my lack of logic, and it was easier to give it a reasonable grade.  (Warning to current students:  do not try this approach with me.)

In today’s Boston Globe, there is a brief article entitled  “You’d sound smarter if you wrote less clearly.” It’s about David Hakes, an economics professor, who simplified a complex argument in mathematical economics.  Hakes explained, ”We managed to reduce the equations in the paper to six. At this stage the paper was perfectly clear and was written at a level so that it could reach a broad audience.”   When the paper was rejected for being “self evident”, Hakes and his co-author decided to make his work less readable and much more complex mathematically, with no added value.  Hakes later wrote ”I personally could no longer understand the paper.”  It was now acceptable to the referees and published.

Hakes’ point extends to Operations Research and other fields that rely on mathematics.    Referees often do not like simplicity because it may make a paper sound “obvious”, and no referee wants to accept a paper that is obvious.  Mathematical complexity usually seems unobvious, and makes the paper more palatable.

As a referee, I have sometimes asked for a substantive revision of a complex paper if I came up with a much simpler version.  For example, I once asked  an 80 page paper to be shortened to around 15 pages, with no loss of content.   But simplifying a paper is, in general, very difficult and is not the task of a referee.  The unfortunate truth is that authors really can increase their odds of publication by making their papers unnecessarily complex.

Japan’s prominent O.R. Ph.D.

As reported in the INFORMS news, the current prime minister of Japan, Yukio Hatoyama,  completed his Ph.D. in Operations Research at Stanford in 1976.   He worked under the supervision of Gerald Lieberman on markov repair models.   His 1977 paper

“Markov maintenance models with control of queue”, Journal of Operations Research Society in Japan 20, 164-181

is available online.

He also co-authored a 1984 book entitled “Stochastic models in reliability theory”.

We all hope that the prime minister is successful in repairing the Japanese economy in a reliable manner.

By the way, I started my Ph.D. at Stanford in the semester after Dr. Hatoyama left the department.  So, I never met him, but wish that I had.

Using Operations Research to improve hospital efficiencies

There is a great article that promotes the use of Operations Research in hospitals that was printed in the Boston Globe this week.  (Actually, it promotes Operations Management rather than O.R.)  Eugene Litvak of Boston University  has worked with Cincinnati Children’s hospital to improve “patient flow” resulting in a far better utilization of their resources.  Litvak’s improvement was comparable to an increase in 100 hospital beds, which in turn was comparable to a savings of $100 million per year.  The hospital was initially resistant to change, as are most hospitals.  But they are very pleased with the results.

Anyone who has experienced a hospital knows that there are often long waits both by patients and staff.   It is an environment that could use lots of O.R. help.  Dr. Michael Long, one of Litvak’s collaborators, put it well.

In some respects, Eugene [a Russian immigrant] was right at home [in dealing with U.S. hospitals].  It reminded him of the way things are done in Russia.

On making games between friends more even

Suppose friends of unequal ability play games on a regular basis.  How can one make the games fair?  A classic approach comes from golf.  Each person subtracts his or her handicap from the score.  In principle this evens up the game.

Here is an alternative approach.  Suppose that A plays B and that A is typically 10 strokes better than B.   So, the first game, the discount for player B is set to 10.  That is, B is permitted to subtract 10 from his score.  If B wins the game, the discount is reset to 9.  If A wins, the discount is reset to 11.   In the long run, A and B win approximately the same number of games, regardless of their relative skills.

This strategy applies to any game that is determined by a final score that is played by friends who want to even the odds.

David Brooks, on why CEOs need O.R.

In today’s New York Times David Brooks wrote an op-ed on the characteristics of good CEOs (“In Praise of Dullness”).

“The traits that correlated most powerfully with success were attention to detail, persistence, efficiency, analytic thoroughness and the ability to work long hours. … What mattered was emotional stability and, most of all, conscientiousness — which means being dependable, making plans and following through on them. …  The second thing the market seems to want from leaders is a relentless and somewhat mind-numbing commitment to incremental efficiency gains.”

I would have preferred it if Brooks had not use words like “dull” and “mind-numbing”.  But I appreciate his implicit support of Operations Research.   Perhaps our field should adopt a new slogan:   “We are Operations Researchers.  We have a relentless commitment to incremental efficiency gains.”

The phrase “and somewhat mind-numbing” is optional.

Saving our newspapers, one electron at a time

Background: Scot Lehigh, an editorial writer for the Boston Globe, asked readers whether an on-line version of the Boston Globe could generate enough revenue to  save the paper. Here is what I wrote to him.

Hi Scott.  Here are some thoughts, random and otherwise.

1. Yes!  I do think that people would be willing to pay for an online newspaper.  I think it could work at a cost of $2 to $3 per week, but this is pure guesswork on my part.
2. The current version of the Boston Globe on line should be kept free.   It needs to be kept free in order to maintain a large list of readers, which strengthens the brand name of the Boston Globe and provides you much more influence locally and nationally.
3. There should be an enhanced version of the Globe that is available on line for electronic subscribers, and free for those who subscribe to the paper.
4. Major newspapers should form a consortium to work out details on enhanced versions.  This is a national problem, and there is no reason that the Boston Globe should be the first to try make it work or should duplicate the efforts of other national newspapers.  It should be a joint effort, possibly supported by schools of journalism as well.

• The enhanced version should  include much more local video.
• The video should be available to all on-line readers, but subscribers could have access to an enhanced resolution video.
• One way of increasing video for very little cost is to encourage readers to submit interesting video.
• You could even have awards for the best submitted videos of the day or week or month or year.
• Include many more high quality photos.  Again, the enhanced version could provide access to higher resolution photos.
• Have an enhanced search engine (perhaps Google or some other company can figure out what would be an improved search engine for newspapers.  The current version is OK but leaves LOTS of room for improvement.)
• Include the comics and puzzles, but possibly at an additional weekly charge of $1 per week.
• Include an enhanced navigation tool for moving from section to section within the newspaper.
• Add some computer intelligence so that articles can be suggested to regular readers based on their reading patterns.
• Create an improved look and feel.  This is extremely hard to get right, but very important.  Check out the Huffington Post for a look-and-feel very well suited for on-line readers.  Unfortunately, it is also a space inefficient way of presenting information to readers.
• Permit access to enhanced versions of other major newspapers, possibly at an extra charge.  (The move to an enhanced version for the Boston Globe would be much more acceptable and much more widely visible if other major papers coordinated on this move, and if subscribers to one newspaper in the consortium received heavy discounts for other papers in the consortium.)
• Include important local news stories from other major newspapers that are part of the “consortium”.
• Charge larger subscription fees for libraries, except for libraries associated with public schools, which should have free access.

Here are two more points.

1. Newspapers have not yet figured out the right advertising model for their on-line version.  But that doesn’t mean that there isn’t a very good model.   One approach is to encourage readers to seek out ads rather than present them to readers whether they want them or not.  Most Boston readers want ads so that they can find bargains.  The on-line version could make it really easy for readers to find discounts and bargains that they are interested in.  The ads should be available in the free and enhanced versions of the on-line newspaper.
2. If the enhanced on-line model is popular and a revenue-generator, you should reconsider issues for the paper, such as advertising, and delivery costs and options.  You should think these through now, but take little action until much later.