SpontaneousSymmetry

Summer conferences are upon us.  Several times a year, members of the high energy experimental physics community gather and share their results with each other.  At these conferences, the major experiments unveil their latest findings to the world, trying to show off the accuracy of their results.  One us conference took place earlier this week in Grenoble, France called the International Europhysics conference on High Energy Physics, or EPS-HEP for short, or even simply EPS.

This particular conference takes place at an extremely interesting time in the history of particle physics.  The Tevatron, the accelerator at Fermi Lab and the former king of high-energy physics, is set to be shut down this Fall.  So, this conference will be one of the last where new data from the Tevatron, mainly from the experiments CDF and D0, will be presented.  At the same time, it’s coming soon after the LHC achieved a milestone by collecting one inverse-femptobarn of data.  As a means of comparison the Tevatron, after about 20 years, has collected 8 inverse femptobarns.  The LHC collected 1 inverse femptobarn since March and is on track to get much more by the end of the year.  So, with this newly produced dataset, made with the record-setting energy of 7 TeV, the LHC experiments have plenty of interesting results.

The most intriguing of these results come from the most up-to-date searches for the Higgs boson.  Both ATLAS and CMS have been working hard to study this elusive particle and are just now getting enough data for those studies to start bearing fruit.  The experiments have many “channels” in which they look for a higgs boson, each consisting of different combinations of measured particles (jets, leptons, missing energy).  Each channel has the potential to discover or put limits on the Higgs, but the real power comes when these channels are statistically combined.

To quickly summarize the results presented:

- Both ATLAS and CMS are now better at searching for the Higgs than the Tevatron, and each individually is able to put stronger limits on the Higgs mass than the Tevatron.

- Both experiments exclude a Higgs boson with about a mass of 150 GeV – 200 GeV

- Neither experiments discovered the Higgs with 5-sigma certainty.

- However, both experiments see a tantalizing excess in the “low” mass range around 120 GeV – 145 GeV, and both see this excess coming from mainly from the Higgs decaying into W boson candidate events.

So, this could be the first taste of finding the Higgs boson. Both experiments will continue to update their results as new data is pouring in at a high rate. Within a year, we should know if the standard model Higgs boson exists with a high level of certainty.

The summaries of the Higgs searches by Atlas and CMS are shown below:

The meaning of the y-axis is “upper limit on the higgs cross-section with 95% confidence.”  Recall that the cross-section is a measurement of how often a Higgs boson is produced.  So, the upper limit on this value represents the maximum value, based on what we measure, that the higgs cross section could have with 95% confidence.  These plots are scaled such that a cross section of 1 represents the standard-model theoretical prediction.  So, when the limit that we measure falls below 1, it means that we have excluded the standard model higgs (at a particular mass point).  The central black line in the middle of the green band represents the “expected limit,” or the limits that we would make if we assume the Standard Model but without a Higgs.  The black dots joined together by a curve represent the measured limit.  So, when reading the plots, you should look for two things.  First, you should ask if the measured limits fall below 1.  Wherever they do, that region is said to be “excluded.”  In addition, one should ask if the measured limits follow the expected limits.  If the measured limits fall within the green band, they are close to expectation within 1 sigma.  If they fall within the yellow, they are close to expectation within 2 sigma.  If they leave the yellow band, they have made a greater than 2 sigma variation from the expected limit.

If the measured limit in a particular region is significantly higher than the expected limit, it could indicate the presence of the Higgs.  People have been very excited about these results for this very reason.  If you look at both ATLAS and CMS’s results, the measured limits are found to be higher than the expected limits in the mass region between about 115 and 140 GeV.  This region particularly interesting because it’s low enough not to be in conflict with results from precision electroweak analyses on the Higgs (which tend to favor a light higgs) but it’s not so low to have been excluded by the direct searches at LEP (the precursor to the LHC which collided electrons and positrons).

The biggest recent news in particle physics came not from the LHC but rather from the Tevatron, the accelerator at FermiLab which, due to budget constraints, is set to be shut down at the end of this year.  The Tevatron has been aggressively perusing the Higgs Boson, trying to find that holy grail of particle physics before relinquishing the spotlight to the accelerator at CERN.  And while the Tevatron has managed put limits on the allowed values of the mass of the Higgs boson, it has thus far been unable to claim discovery.

However, the Tevatron may have stumbled upon some extremely exciting new physics that would demand new theories of fundamental particles.  The CDF experiment, one of the experiments of the Tevatron, published a paper recently showing a “bump” in the invariant mass spectrum of events involving W Bosons.  This feature is not present in Standard Model simulations and therefore could be a tell-tale sign of “new physics.”  On the other hand, it could simply turn out to be a feature of detector resolution.  So, let’s take some time to understand what exactly the CDF experiment saw and to think about how significant a discovery it is.

This “bump” was found in a search of what are called “diboson” events, which are events that involve two W bosons, two Z bosons, or and W and a Z boson.  (The term boson refers to a particle that carries a force, such as the W or Z, which mediate the weak force.  The photon too is a boson and it mediates the electro-magnetic force.  The gluon is the boson which mediates the strong force, and the Higgs is a yet-to-be-discovered boson which, in short, creates mass).

In the diboson events that were studied in this analysis, one of the bosons decays into a lepton (meaning an electron or it’s cousin the muon) and the other decays into two quarks.  These quarks eventually become jets, which are messy cones of particles that come from the effects of quantum-chromodynamics (QCD).  In addition, these events contain missing energy that comes from invisible neutrinos escaping the detector.  So, experimentalists collected these events from CDF and calculated the “invariant mass” of the two jets.  Since these jets came from one of the bosons (either the W or the Z), one would expect this mass to be close to the W or the Z mass, which are 80 and 91 GeV, respectively.  In practice, since these jets are so messy, the resolution of the detector isn’t fine enough to distinguish between those two masses, so one expects to see one giant mass peak between 80 and 90 GeV.  When experimentalists looked at these events, this is indeed what they saw:

The red in the plot above represent these diboson events that the experimentalists were looking for.  The different colored histograms are background, mostly coming from simulation.  The black dots with the small vertical bars are the data that was actually measured by the experiment.  So, if the sum of the colored blobs add up to the black data points, then everything is well-understood and agrees with standard model prediction.  And this is the case, for the most part.  One can clearly see that the black lines follow the red bump where simulation says it should be, which essentially means that experimentalists have found the diboson events they were looking for.  At the same time, there is a lot of background.  In this analysis, there is much more background than signal, moss of which comes from “W plus jets” events in green.

So, what’s all the excitement about?  The Diboson signal (red) indeed behaves as it should.  But there is another part of the plot that stands out.  If you look at the falling slope at around 140 GeV, you’ll see that the measured data (the black bars) are a bit higher than they should be (the data should match the sum of all the colors, including green and the little red on top).  For a few bins in a row, there are more events measured in the data than one would expect from the standard model simulation.  Many believe this excess of events is evidence for new physics at work, and perhaps is the calling-card of a new yet-undiscovered particle.

If you take the data and subtract everything but the red, you’ll end up with this plot:

There is a first bump at around 90 GeV where we expect the standard diboson signal to be.  However, the excess around 140 GeV is clearer as a bump (the red and blue lines are fits to these shapes and have been included to guide the eye).

The immediate objection of many is that, looking at the original plot without the background subtraction, the excess of events seems very small.  Indeed it is, there is a lot of background from W+Jets, and this new signal is just a small bump on top of that giant hill.  And the bottom line is the following: if one doesn’t understand the hill extremely well, then one can’t really determine if the bump is real or not.  Is there a new physics signal, or is our simulation of the backgrounds just slightly wrong?  This, in essence, is the name of the game.  Particle physicists are mostly concerned with determining how well they understand their backgrounds and what type of statistical statements can be made using that understanding.

For example, this excess of events is seen in the distribution of the mass of jets in diboson events.  Therefore, in order to make a strong statement using this distribution, one must have a very good understanding of how well these jets have been measured.  Jets are very messy objects and involve a lot of particles moving in a large cone.  It is difficult to measure the energy of all these particles and to be sure that some energy isn’t missed or over-counted.  Some have suggested that the bump seen merely comes from a misunderstanding of the energy of these jets.  One experimentalist working on the CMS collaboration demonstrated in a simple animation how this effect could work (taken from Quantum Diaries Survivor:

In this (somewhat crude but very accessible) animation, the energy of the jets, and therefore their mass distribution, is varied.  One can see that when the energy is scaled upwards by up to 7%, the excess around 140 GeV seems to disappear (and this is especially noticeable when looking at the top plot showing the difference between signal and background).  While this is far from definitive, it suggests the types of studies that must be performed before this hint of new physics can be accepted or ruled out.  And while the validity of this claim is far from certain, it is very exciting for a particle physicist and is hopefully but a small taste of the plethora of new physics that will be discovered in the next decade by the LHC or the many other experiments around the world.

Recently, the New York Times released an interactive version of the game Rock Paper Scissors.

RPS is essentially a break-even game. By playing Rocks, Paper, or Scissors randomly, one is guaranteed on average to win as often as one’s opponent. Therefore, the only type advantage that one player can have over another would be psychological. In essence, can beat a human opponent by taking advantage of people’s poor ability to generate random numbers. A person, in attempting to be random, or even in an attempt ing to gain a psychological advantage, may fall into patterns that can be exploited.

The New York Times’ website designed such a program to exploit these tendencies in human opponents. The designers of the Times’ interactive game claim to use the past history of a person’s moves to anticipate and counter that individual’s next move. The article explained in general how this works. The opponent computer algorithm builds up in memory a string of a person’s move history as he plays the game. It then searches through the person’s last few moves and looks for a matching pattern earlier in the game. Based on what the person did after that earlier matching patter, the computer guesses what the person will do as his next move.

For example, let’s say that after 10 games a person’s history looked like RRSRPSRSRP. The last three moves that this person made were “SRP”. The algorithm would look through his history for another time that he threw “SRP” in that order: RRSRPSRSRP. Beginning at the third move, the person threw “SRP” followed by an “S.” Therefore, the algorithm would guess that the person is again going to throw an S and, to exploit this potential pattern, the computer would counter with a R (since Rock beats Scissors). The algorithm starts by looking for long strings 5 moves long and then searches for shorter strings if no longer string patterns are matched.

This algorithm is pretty simple, but apparently somewhat effective. However, since it’s a fixed algorithm, it’s perfectly exploitable. Simply put, if one were able to perfectly recreate the algorithm based one ones own moves, one could predict the computer’s guess for the human’s move. With this guess, one knows what the computer will throw, and therefore one can simply throw the hand that beat’s the computer’s move. It’s simple and quite insidious. So, I set out to do just that. I wrote a short python script that, as best as it was described by the Times, copies the Times’ algorithm. By knowing the algorithm, I can do my best to beat the machine at it’s own game. If the computer thinks that I’ll throw a Rock, it will throw a Paper. But if I know that the computer think I’ll throw a Rock, I know it’ll throw a paper, and instead I can throw a Scissors.

My Code: RocksPaper

I’ve included my python code and an example of how to run it. I chose python because itsinteractive interpreter allows me to use the code to in real time play against the computer. An example of a round I played is as follows: After the 8th round:

>>> Game.playNext()
Throws so far: ['R', 'R', 'P', 'S', 'S', 'R', 'S', 'S']
Last 4 : ['S', 'R', 'S', 'S']
Matching Plays: []
Last 3 : ['R', 'S', 'S']
Matching Plays: []
Last 2 : ['S', 'S']
Matching Plays: ['R']
Suggested Play: S

As you can see, MY algorithm started by looking at my last 4 throws and seeing if there were any such throws in my history. It found none, and then moved on to 3 and finally 2. My last two throws were “S,S”. My code found that I had previously thrown SS starting on the 4th round, and after that pair, I had thrown a R. The Times’ program, based on that, would anticipate that I would throw a R, so it throws a P. But since I know it’s going to throw a P, I throw a S.

Using this strategy, I overwhelmingly defeated the computer. It was pretty satisfying, actually. I felt a little bad seeing the computer be so wrong so often. But I realized that I was simply doing what the computer was attempting to do to me. The algorithm was attempting to understand the complexities of human psychology and use them against us. I was simply using the simplicity of the computer’s psychology against it.

After 57 games (when I arbitrarily stopped playing), I had amassed 40 wins to the computer’s 6, with 11 ties.

As the complete nerd that I am, I wanted to quantify my victory, so I decided to compare that level of victory to the amount of victories that one would expect using a random-only strategy. In Stat-speak, I wanted to calculate the p-value of a random RPS strategy.

It’s easy to show that the probability of a particular end-game result of W wins, T ties, and L loses (ignoring order, which isn’t important) is given by:

(The brackets indicate the binomial, or “choose,” function).

One can then simply preform a sum to calculate the total probability of obtaining as many or more wins minus losses from random chance. Using my numbers, I calculate that the chance of a random strategy doing as well or better as my strategy is 9.26 * 10 ^ -9. This is certainly a statistically significant result (it’s better than the 5 sigma significance required by higher energy physics, which corresponds to a p value of 5.7 * 10^-7. My value falls somewhere between 5 and 6 sigma.)

Though the problem is pretty trivial in terms of computer science, I’m glad that the NYTimes made this demonstration. I imagine that the authors of this program were inspired to the success of Watson in playing Jeopardy. I think it’s a good thing that the Times is seeking to demonstrate how powerful even very simple computer algorithms can be. And it was certainly a fun exercise to try to beat the Times’ model. If I were making my own model, I would make a few changes (for example, I would consider simultaneously the result of different size string searches). But the point wasn’t to make the best possible computer RPS player. Rather, it was to encourage people to think algorithmically and to elucidate how programmers approach computer challenges.

Supersymmetry is one of the most popular theories in theoretical physics.  Really, it’s a large set of theories which are used to address, among other things, what is known as the Hierarchy problem.  The Hierarchy problem, in a nutshell, is the issue of why there appear to be two very different scales in the universe.  The energy scale of W, Z and Higgs bosons (collectively known as the “Electroweak” scale) is much smaller than the energy scale of gravity (the “Plank” scale): the masses of W, Z, and Higgs bosons are all around 100 GeV, where as the intrinsic energy scale of gravity is around 10^19 GeV.  In our current understanding of physics, this difference seems to come about due to almost miraculous fine tuning in physical equations.  Such perfection seems unnatural in the way it works together.  Generally, physicists don’t prefer that their physical equations rely on arbitrary coincidence to work.  Instead, theories like Supersymmetry introduce new particles which instead lead to a natural and beautiful solution to the hierarchy problem.

Because the theory of Supersymmetry introduces a wide spectrum of new particles, it is directly testable in experiments like ATLAS at the LHC.  Of course, other accelerators have been searching for Supersymmetry (often shortened to SUSY) for some time, and no one has yet been able to discover any SUSY particles.  When a physicist fails to make discovery, he can instead set limits.  He can essentially say, “If SUSY of this form or that existed, we would have seen it in this experiment.  Since we didn’t, we can now restrict the possible forms of SUSY that could potentially exist.”  This is all we’ve been able to do thus far.  Recently, using 2010 data, experiments at the LHC have managed to set the strongest limits yet on SUSY theories.

Before we get into that, let’s talk a bit about how one looks for SUSY at a particle accelerator.  Well, one does that is essentially the same way that one looks for any other physical signature: one determines how that signature would look, what particles would emerge from it, how they would interact with the detector, and what properties those particles would have.  One also considers other physical processes that, by looking similar, can mimic the physical process of interest, and one attempts to measure how much of these “backgrounds” would be present.

To be very concrete, let’s look at a plot released by a recent ATLAS paper on SUSY:

This plot shows several things at once.  The x-axis is a quantity which this paper calls the “effective mass,” which here is a combination of the energy of jets, leptons, and missing energy of a collision event.  It isn’t strictly necessary to understand the exact nature of this quantity to interpret this plot.  The y-axis is the number of events measured or that one would expect to see in a particular bin (notice the logarithmic scale).  Each of different colored stacks represents the expected distribution of a some background (most of these distributions come from simulation).  For example, the dark and light green stacks come from top quark events, the dark brown comes from a class of processes labeled “Diboson,” and the small amount of blue comes from W-bosons.  The dotted white and black lines are where SUSY would be if it were present.  Finally, the solid black dots are the actual data: they’re what was actually measured by the detector.  If SUSY existed, one may expect there to be black dots out by the SUSY curve toward the right.  However, we see no such dots (it should be noted that, if you look at the y-axis, the expected number of events is less than 1 in that region, so it’s possible for SUSY to exist but not show up on this graph.  This is a somewhat low-statistics analysis).  Instead, you do see the black dots at the top of the stack of backgrounds, where we expect them to be, where they are predicted to be by the Standard Model.

To evaluate how strongly the experiment’s data supports or excludes a SUSY theory, experimentalists construct statistical models to describe various potential theories and determine which of these models is closest to the observed data.  Of course, I’m glossing over what I think is one of the most interesting parts of the discovery process, but I would need a much bigger space to accurately describe the statistical techniques used at ATLAS.  But using the observed data and some nice statistics, ATLAS, as well as its rival experiment CMS, have managed to put very strong limits on how SUSY must look, if it exists at all.

This plot shows these newly placed limits.  As I said earlier, SUSY is really a collection of theories, and many of these are parameterized by some standard quantities.  Each point in this plot represents one particular theory in a continuum of theories, each with different values of two x-axis and y-axis parameters (again, their exact meaning isn’t necessary to describe here).  The colored blobs on the bottom-left of the plot are limits that have already been placed by earlier experiments.  Areas shaded in those blobs are “excluded,” meaning that a SUSY theory with those parameters have been experimentally rejected (with a 95% confidence level).  The red line represents the NEW limits from ATLAS, and the black line is the limit placed by CMS.  Areas below and to the left of those lines have been excluded by these experiments (here, ATLAS places a slightly stronger limit than CMS).  The other blue lines describe what limits one would EXPECT to observe and have been included to give context to the actual OBSERVED limits.

These limits only apply to a small subset of all possible SUSY theories.  In reality, SUSY has tens of free parameters, and it would be impossible to draw on a piece of paper the 30 or so dimensional space of SUSY models that have been excluded by this new experimental data.  This is unfortunate, particularly because it means that many people think of SUSY in a narrow way since, historically, only a few parameters or a few particular models appear in papers.  This is arbitrary, and there’s no real reason that SUSY should conform to the taste or historical tradition of physicists.  Instead, to describe physical results in a more general way, experimentalists should strive to break out from the tradition of publishing only a small subset of the excluded space.  In the example of SUSY, it would be much preferable to instead publishing the entire, 30-dimensional space of excluded theories.  Again, it’s impossible to draw this on paper, but it’s the 21st century, and the concept of publishing should conform to the world in which we live.  Physicists could release the excluded space as some persistent data-structure in one form or another that a person could project into any subspace he desired for the sake of visualization.  So, if a theorist weren’t interested in the two parameters that were published on the arxiv paper, he could instead view the excluded hyperspace in any 1, 2, or 3 dimensional space of SUSY parameters.  It’s time that physics publications catch up to modern times.

Last night the New York Knicks made a franchise-changing trade when they essentially swapped several players for Carmelo Anthony, a star who has long stated his desire to play in New York.  It was a polarizing trade that many have strong opinions about.

Nate Silver of 538 fame made a simple analysis arguing why he believes the trade did not help the Knicks as much as some believe.  The central axiom to Silver’s discussion is that the goal of an NBA general manager is to:

1. Acquire players who produce above-average value relative to the salaries they are making.

2. Exploit the loopholes in the salary cap so that you spend more money than other teams.

To illustrate his first axiom, Silver used an algorithm developed created by John Hollinger to determine how many wins each player “adds” to a team.  Silver described those players who add many games with a small salary as high in value.  Silver argued that, using his metric, Anthony skill to value ratio isn’t as high as that of other superstars around the league, and his large salary will end up hindering the Knicks’ future moves and progress.

To build upon Silver’s arguments, one can see the distribution of the Knicks players in the Salary-Wins plane both before and after the trade:

Distribution before the trade                           Distribution after the trade

If one were to strictly interpret Silver’s method, one would see that the vast majority of wins come from a few players, all of which have a pretty similar ratio of wins to salary.  The rest of the players add relatively few wins at the cost of a small amount of salary.  Before the trade, the Knicks had several players in a middle range in terms of both wins added and salary cost.  Post trade, the Knicks are a much more polarized team, with few high-win players and the rest clustered close to zero in terms of salary and wins.

Yet, simply by looking at these plots, it becomes pretty clear that Silver’s model is vastly oversimplified.  It treats the majority of players on a team as small line segments only incrementally advancing the team’s wins.  The reason for this model’s failure is that it completely factorizes the players on the team: it treats a team as a blind sum of individuals instead of as a collect effort.  But anyone who knows basketball, or any sport, knows that this is far from true.  Players either enhance or hinder one another.  A team consisting of only Carmelo Anthony and four middle-schoolers would not win 14 games against an NBA team because Anthony requires a supporting cast to succeed.  And the effects of this cast isn’t easily quantifiable in terms of numbers, wins, or salaries.

In sports, there has emerged a false-dichotomy between “statistics” people and those who understand a game using their “instincts.”  But too often, using statistics in sports means using massive simplifications and trivial models.  It means giving too much weight to those things that are easily measurable.  And it too often means treating a team as a sum of individuals.  While I don’t mean to discredit the efforts of the many who have gone a long way to adding legitimacy to sports statistics, I believe the usefulness of those derived numbers models is limited.  One should not take them as law.  Sports statisticians should do they best they can with the tools and data available, but having created a model or statistic, they should be aware and open to its limitations.

Last week the first of two games of Jeopardy in which the IMB supercomputer Watson faces off against human opponents was played.  And, in the end, the computer asserted its dominance with a commanding victory.

Over the past three days, IBM’s custom supercomputer Watson off against against humans in two televised games of Jeopardy.  The challenge of designing and programming a computer to win at Jeopardy is the successor to IMB’s Deep Blue, which faced off against chess grand master Garry Kasparov.  I will admit that, watching the show, I couldn’t help but root for Ken Jennings and Brad Rutter.  While I have been excited about Watson for a long time, watching the match, I realized that I have a soft spot for humans.

So, what does Watson’s victory last night mean in the greater scheme?  Is this the beginning of a revolution in computing?  Is this the beginning of a more literal revolution where computers rise up and take over our planet?  I believe that it’s neither.  While many aspects of Watson are impressive, I found myself a bit underwhelmed by what he showed.  Perhaps he (it?) was too convincing, too good at what he did for me to remain impressed.  As I was watching, I found myself thinking it obvious that a computer would be able to answer many of the questions asked.  At the same time, I knew that my reflex-like skepticism was  ignoring the difficulty in parsing human language, which is the real achievement of Watson.

The challenge that Watson represents is not about storage or search.  These are the things that computers are inarguably incredibly good at, and we have all become jaded to a machine’s dominance.  Retrieving a well-defined fact is simple if it is stored in a well organized database and the question is merely a query of that database.  This is a trivial task in computer science.

The real challenge is deciphering what a question is looking for and what type of answer is required.  In many ways, Watson is an English grammar machine.  It has to break down the sentence structure of the input question (or “answer,” in Jeopardy lingo) and decide what object it’s looking for.  Does a question require a date in time, a person, the name of a song?  Presumably (Watson’s exact algorithms are kept a secret) the computer, once it identifies the type of thing being looked for, evaluates out of many possibilities those objects of that type that are most associated with the key words appearing in the question.  Watson actually uses more than a thousand different algorithms to come up with these possible answers, and its level of confidence is derived from how many algorithms return the same answer.

Using this sort of structure, some Jeopardy questions will be inherently easier than others.  Based on his performance over the last few days, it was clear that Watson was very good at questions questions where the question essentially gives one piece of key information that is unambiguously associated with the answer.  If asked who composed a particular symphony, Watson would be very likely to come up with the answer.  It must realize that it’s looking for a person and decide the person most associated with that symphony, which would easily be the composer.

Watson struggled more where questions gave information from two directions and were more implicit in their suggestions.  One question read, “Stylish elegance, or students who all graduated in the same year.”  This sort of question doesn’t have a one-to-one nature to it.  Rather, it describes two different partial definitions of a word.  Watson incorrectly answered “chic,” which was probably highly associated with “Stylish elegance,” but didn’t match that well with the “students who graduated” part of the question.  The real answer is “class.”

Similarly, Watson failed to answer on of the Final Jeopardy questions correctly.  The question, “Its largest airport was named for a World War II hero; its second largest, for a World War II battle” is extremely complicated.  There aren’t really any key words that jump out.  Watson, using its many algorithms, had to try to associate “city” with “World War II” and “airport.”  While the exact relationship between the two airports and how they relate to World War II is obvious to a human, there’s no reason to believe that Watson was really able to understand that relationship in a meaningful way.  Watson could only search those concepts and try to find the city with the closest tie to those ideas.  It answered, “Toronto,” which isn’t actually a US city as the category required.  The correct answer is “Chicago,” which both humans were able to answer.

Some have argued that Watson’s real advantage is it’s ability to push the answer button faster and more reliably than the humans.  I think there’s a lot of merit to this argument.  Watson is fed the question as soon as it appears on the screen, when Alex starts to read it.  In Jeopardy, contestants can only buzz in after the host has finished reading the question.  Watson is electronically told when it can start to answer a question.  As far as I understand, the same signal that activates the buzzers also tells Watson that it is now allowed to buzz in.  If it has calculated an answer by that point, it will use it’s lighting-fast  hydraulics to buzz in almost immediately, faster than any human can.  So, the way the game is set up, Watson essentially has first dibs on any question that it knows by the time Alex finishes reading.  And, as any Jeopardy contestant will tell you, the buzzer is the key to success on the show.

So, how should we feel about Watson?  I think we should be nothing but excited about success that Watson has shown, but especially for its future applications.  Watson is really a model for an interface between a real human question and a huge dataset.  And, while it’s not perfect at a game like Jeopardy, it’s pretty darn good.  A Watson-like system would be much better in an environment where it’s not trying to be tricked, where the human asking the question really wants Watson to come up with the right answer.  I’m think a Watson system would be particularly invaluable in the medical community.  In a field with an exponentially exploding amount of research and data on medicine and health, a tool which can answer questions phrased in plain english would be extraordinarily helpful.  Watson isn’t there yet, and it would have to be tailored for the particular needs of any non-Jeopardy practical application.  But Watson it’s a solution, it’s a spectacle.  It’s an example of what computers can do.  It’s goal isn’t to be the end but rather the beginning.  Watson is a small, self-contained demonstration of what future computers will be.  It is pointing where research should aim and taking a major step in that direction.

The LHC has just made a major announcement about it’s plan for the next several years.  The various directors of the Large Hadron Collider and its experiments have just announced that the LHC will run in physics mode through the end of 2012 instead of only through 2011 as was originally planned.  This schedule is a major change from the most recent plans in which the accelerator was to run only until the end of this year and then shut down for a year of repairs and upgrades.  Now the plan is for the LHC to go through a nearly two year long shutdown starting at the end of 2012.  This announcement comes on the heels of the announcement that the Tevatron, the 1 TeV particle collider at Fermilab in Illinois, will cease operations in September of 2011.

The plan is to continue to collide protons at a center-of-mass energy of 7 TeV until the end of 2012, which is the same energy that was used throughout 2010.  Initially, there was some discussion of increasing the energy to 8 Tev (which is still short of the design energy of 14 TeV).  However, to ensure stability and to avoid any glitches that may have come from an increase in energy (or any outright disasters similar to the September 2008 incident), an increase in energy, for now, is no longer a part of the plan.

This move significantly alters the physics prospects of the LHC over the next few years.  The biggest question is how these changes effect the LHC’s prospects of discovering the Higgs Boson.  Many believed that the Tevatron would have had a good chance at finding the Higgs with high significance if allowed to run for a few more years.  But the recent announcement of it’s cancellation means that the only hope for finding the elusive particle lies in Europe (there is a joke among physicists that Bosons can only be discovered in Europe and fermions are discovered in the US, which at least historically has seemingly been the case).  Over the next two years it is estimated that the ATLAS detector will collect 5 Inverse femptobarns worth of data.  At 7 TeV, this MAY be enough statistics to officially discover the Higgs (by having 5-sigma significance), or at least to have a good idea of its mass.  However, inn order to maximize the chances of making discovery, analysts will have to become more aggressive in their strategies for seeking out the Higgs.

A move to 8 TeV would not simply be symbolic.  The small increase in energy could mean a large increase in the amount of Higgs produced and could lead to much better statistics (depending on the Higgs’ mass).  The decision to not increase the collider’s energy dramatically effects the prospects for potential discoveries other than the Higgs.  It essentially means that supersymmetry (if it exists) will not be discovered before 2015 or so.

I think the LHC is doing the right thing by extending this year’s upcoming run (to resume again in a month or so) by another year.  It will reduce the chances of an accident and ensure that a lot of good data comes out of the machine as quickly as possible.  It will hopefully (with luck and hard work) lead to an earlier discovery of the Higgs (there is no doubt that during the two year break after 2012, the data will be reprocessed and reanalyzed as much as possible to find any hints of new physics).  And while it will delay some interesting physics, it will ensure that the machine is well understood and working well as safely and as quickly as possible.

Ever since I was a kid, I was interested in the concept of using evolution to program and solve problems. I didn’t really know what it was called at the time or if it was even feasible, but I was fascinated with the idea of having a computer learn to program itself, to adapt, and evolve (this probably grew out of reading too many Michael Crichton novels). This idea is the basis of a very real and quite useful set of programming techniques called “genetic algorithms.”

The goal of genetic programming is to have a computer program “evolve” by gradually adjusting its own parameters based on how well it preforms. There are many different implementations of Genetic Programming, but this is typically done by creating many different versions of a program, all with different parameters, and seeing which version preforms the task best. Those programs which preform the task best are combined together in some way to create a new or several new offspring programs (mimicking reproduction). In addition, genetic algorithms often include some level of mutation in which aspects of a program randomly change. So, one starts out with a collection of initial programs and, by combing and mutating these programs, one creates a next “generation” of these programs. Those that work best continue on, and those programs that fail aren’t recombined into a next generation. It’s controlled Darwinism.

I came across a fun example of this concept:

boxcar2d.com

This site, before your eyes, uses genetic programming to create a simple version of a car that can travel along a bumpy path. The program creates cars out of triangles, wheels, and axels, and sees which cars can drive the furthest.

Those cars that make it the farthest are recombined into newer cars (with some additional mutations). Those that outright fail are lost (they go extinct). This simple game is a good demonstration of how structure and design can emerge spontaneously out of a simple set of rules. Initially, most of the cars are odd shapes with wheels sticking out that may or may not even touch the ground. Some simply fall and fail to even move. Others may crawl a few feet if their wheel happens to be in the right position.

Those that can move a bit breed with each other and form a new generation, which should be slightly better at driving. This process continues as long as you’re willing to keep your browser open, and, as one would expect, with each generation the cars are able to drive further.

In addition, they become more “car-like.” They learn that having two wheels, one at either end, makes them more stable and helps them survive the bumpy road. So, as time goes on, they tend toward the expected shape of a vehicle (remember, this happens with no input into the program as to what a car is or should look like).

However, other features that one would not necessarily expect also emerge (at least, they did during my time as evolution’s third-party observer). Because the program uses a somewhat realistic physics engine that determines how the cars bounce and rotate when traversing the track, the balance and weight distributions of the cars become important. A common feature that I saw was the presence of an angular, forward-pointing weight on top of the car which, as far as I could tell, helped the vehicle stay horizontal when going over jumps. If designing a car by hand, one may not have come up with this sort of structure. But, because the car optimizes itself for a particular set of conditions, even without any external, intelligent design, it can seek out optimal solutions (as long as those solutions can be reached as the aggregate of small, local changes).