Skip to content

Stopped Watches and Data Analytics

May 23, 2017

Is this a new or old paradox?

UK Independent source—and “a gentle irony”

Roger Bannister is a British neurologist. He received the first Lifetime Achievement Award from the American Academy for Neurology in 2005. Besides his extensive research and many papers in neurology, his 25 years of revising and expanding the bellwether text Clinical Neurology culminated in being added as co-author. Oh by the way, he is that Bannister who was the first person timed under 4:00 in a mile race.

Today I cover another case of “Big Data Blues” that has surfaced in my chess work, using a race-timing analogy to make it general.

Sir Roger also served as Head of Pembroke College, one of the constituents of Oxford University. He was one of three august Rogers with whom I interacted about IBM PC-XT computers when the machines were installed at Oxford in 1984–1985. Sir Roger Penrose was among trustees of the Mathematical Institute’s Prizes Fund who granted support for my installation of an XT-based mathematical typesetting system there, a story I’ve told here. Roger Highfield and his secretary used an XT in my college’s office, and I was frequently called in to troubleshoot. While drafting this post last month, I received a mailing from Sir Martin Taylor that Dr. Highfield had just passed away—from his obit one can see that he, too, received admission to a royal order.

Dr. Bannister was interested in purchasing several XTs for scientific as well as general purposes at Pembroke. At the time, numerical performance required purchasing a co-processor chip, adding almost $1,000 to what was already a large outlay per machine by today’s standards. I wish I’d thought to say in a quick deadpan voice, “let it run four minutes and it will give you a mile of data.” (Instead, I think the 1954 race never came up in our conversation.) Today, however, data outruns us all. How to keep control of the pace is our topic.

Roger Bannister 50-year commemorative coin. Royal Mint source.

Stopped Watches Can Be Correct

As shown above in the commemorative coin’s design, the historic 3:59.4 time was recorded on stopwatches. We’ll stay with this older timing technology for our example.

Suppose you have a field of 200 milers. Suppose you also have a box of 50 stopwatches. For each runner you pick a stopwatch at random and measure his/her time. You get results that closely match the histogram of times that were recorded for the same runners in trials the previous day.

How good is this? You can be satisfied that the box of watches does not have a systematic tendency to be slow or to be fast for runners at that mix of levels. Projections based on such fields are valid.

The rub, however, is that you could have gotten your nice fit even if each individual watch is broken and always returns the same time. Suppose your field included Bannister, John Landy, Jim Ryun, and Sebastian Coe, with each in his prime. They would probably average close to 3:55. Hence if one of the 50 watches is stuck on 3:55, it will fit them well. It doesn’t matter if you actually draw the watch when measuring the last-place finisher. The point is that you expect to draw the watch 4 times overall and are fitting an aggregate.

Indeed, you only need the distribution of the (stopped) watches to match the distribution of the runners under random draws. You may measure a close fit not only in the quantiles but also the higher moments, which is as good as it gets. Your model may still work fine on tomorrow’s batch of runners. But at the non-aggregate level, what it did in projecting an individual runner was vapid.

Another Quick Example

Here is a hypothetical example in the predictive analytic domain of my chess model. Consider a model used by a home insurance company to judge the probabilities of damage by earth movement, wind, fire, or flood and price policies accordingly. I’ve seen policies with grainy risk-scale levels that apply to several hundred homes in a given area at one time. The company only needs good performance on such aggregates to earn its profit.

But suppose the model were fine-grained enough to project probabilities on individual homes. And suppose it did the following:

  • Home 1: Earth 0.047, wind 0.001, fire 0.001, water 0.001;    total risk 0.05.

  • Home 2: Earth 0.001, wind 0.047, fire 0.001, water 0.001;    total risk 0.05.

  • Home 3: Earth 0.001, wind 0.001, fire 0.047, water 0.001;    total risk 0.05.

  • Home 4: Earth 0.001, wind 0.001, fire 0.001, water 0.047;    total risk 0.05.

This is weird but might not be bad. If the risks average out over several hundred homes, a model like this might perform well—despite the consternation homeowners would feel if they ever saw such individual projections.

Of course, “real” models don’t do this—or do they? The expansion of my chess model which I described last Election Day has started doing this. It fixates on some moves but gives near-zero probability to others—even ones that were played—while giving fits 5–50x sharper than before. If you’ve already had experience with behavior like the above, please feel welcome to jump to the end and let us know in comments. But to see what lessons to learn from how this happens in my new model, here are details…

How My Original Model Works

My chess model assigns a probability {p_{i,t}} to every possible move {m_i} at every game turn {t}, based only on the values {v_i = v_{i,t}} given to those moves by strong computer chess programs and parameters {s,c,\dots} denoting the skill profile of a formal player {P = P(s,c,\dots)}. The programs list move options in order of value {v_1 \geq v_2 \geq \cdots v_i \cdots} for the player to move, so that {v_1 - v_i} is the raw inferiority of {m_i} in chess-specific centipawn units.

The model asserts that the parameters can be used to compute dimensionless inferiority values {z_i}, from which projected probabilities {p_i} are obtained without further reference to either parameters or data. The old model starts with a function {x_i = \delta(v_1,v_i)} that scales down the raw difference according to the overall position value {v_1}. Then it defines

\displaystyle  z_i = \left(\frac{x_i}{s}\right)^c .

Lower {s} and higher {c} both decrease the probability {p_i} of playing a sub-optimal move {m_i} by dint of driving {z_i} higher. The effect of {s} is greatest when {x_i} is low, so {s} is interpreted as the player’s “sensitivity” to small differences in value, whereas {c} governs the frequency of large mistakes and hence is called “consistency.” My conversion represents each {p_i} as a power of the best-move probability {p_1}, namely solving the equations

\displaystyle  p_i = p_1^{\exp(z_i)};\qquad p_1 + \cdots + p_\ell = 1,

where {\ell} is the number of legal moves in the position. The double exponential looks surprising but can be broken down by regarding {u_i = \exp(-z_i)} as a “utility share” {u_i} expressed in proportion to the best move’s utility {u_1 = 1}, then {p_i = p_1^{1/u_i}}. Alternate formulations can define {u_i} directly from {x_i} and the parameters, e.g. by {u_i = \frac{1}{(1 + x_i/s)^c}}, and/or simply normalize the shares by {p_i = p_1 u_i = u_i/(\sum_j u_j)} rather than use powers, but they seem not to work as well.

This “inner loop” defines {P(s,c,\dots)} as a probability ensemble {t \rightarrow \{p_{i,t}\}} given any point {(s,c,\dots)} in the parameter space. The “outer loop” of regression needs to determine the {P(s,c)} that best conforms to the given data sample. The {\{p_{i,t}\}} determine projections for the frequency of “matching” the computer’s first move and the “average scaled difference” of the played moves {m_{i_t,t}} by:

  • {\mathsf{MM}_p = \frac{1}{T}\sum_{t=1}^T p_{1,t}}.

  • {\mathsf{ASD}_p = \frac{1}{T}\sum_{t=1}^T \sum_i p_{i,t}\delta(v_{1,t},v_{i,t})}.

The regression makes these into unbiased estimators by matching them to the actual values {\mathsf{MM}_a} and {\mathsf{ASD}_a} in the sample. We can view this as minimizing the least-squares “fitness function”

\displaystyle  F(s,c) = w_1 (\mathsf{MM}_p - \mathsf{MM}_a)^2 + w_2 (\mathsf{ASD}_p - \mathsf{ASD}_a)^2,

where the weights {w_j} on the individual tests are fixed ad-lib. In fact, my old model virtually always gets {F(s,c) = 0}, thus solving two equations in two unknowns. Myriad alternative fitness functions {F'(s,c)} using other statistical tests and weights help to judge the larger quality of the fit and cross-validate the results.

In my original model, all is good. My training sets for a wide spectrum of Elo ratings {E} yield best-fit values {(s_E,c_E)} that not only give a fine linear fit {\hat{E} = \alpha + \beta s + \gamma c} with residuals {|E - \hat{E}|} small across the spectrum, but the individual sequences {[s_E]} and {[c_E]} also give good linear fits to {E}. Moreover, for all {E} and positions {t} the projected probabilities {\hat{p}_{i,t;E}} derived from {(s_E,c_E)} have magnitudes that spread out over the reasonable moves {m_i}.

The New Model

My old model is however completely monotone in this sense: The best move(s) always have the highest {\hat{p}_{i,t;E}}, regardless of {E}. Moreover, an uptick in the value {v_i} of any move {m_i} increases {\hat{p}_{i,t;E}} for every {E}. This runs counter to the natural idea that weaker players prefer weaker moves.

The new model postulates a mechanism by which weaker moves {m_i} may be preferred by dint of looking better at earlier stages of the search. A new measure {\rho_i} called “swing” is positive for moves {m_i} whose high worth emerges only late in the search, and negative for moves that look attractive early on but end with subpar values. The latter moves might be “traps” set by a canny opponent, such as the pivotal example from the 2008 world championship match discussed here.

A player’s susceptibility to “swing” is modeled by a new parameter called {h} for “heave” as I described last November. The basic idea is that {v'_i = v_i - h\rho_i} represents the “subjective value” of the move {m_i}, so that {\delta(v_1,v_i) + h(\rho_i - \rho_1)} represents the subjective difference in value. The idea I actually use applies swing to adjust the inferiority measure:

\displaystyle  z'_i = z_i + \left(\frac{h(\rho_i - \rho_1)}{s}\right)^{ac} = \left(\frac{\delta(v_1,v_i)}{s}\right)^c + \left(\frac{h(\rho_i - \rho_1)}{s}\right)^{ac},

where {a} is a fourth parameter and {x^{ac}} for negative {x} is defined to be {-|x|^{ac}}. Dropping {s} from the second term and raising it to just the {a} not {ac} power would be mathematically equivalent, but coupling the parameters makes it easier to try constraining {h} and/or {a}. (In fact, I’ve tried various other combinations and tweaks to the formulas for {\rho_i} and {z'_i}, plus four other parameters kept frozen to default values in examples here. None so far has changed the picture described here.)

Note that the formulas for {z'_i} preserve the property {z'_1 = 0} for the first-listed move {m_1}. When {m_2} has equal-optimal value, that is {v_2 = v_1}, {\rho_2} cannot be negative and {\rho_2 - \rho_1} is usually positive. That makes {z_2 > 0} and hence reduces the share {u_2} compared to {u_1}. The first big win for the new model is that it naturally handles a puzzling phenomenon I identified years ago, for which my old model makes an ad-hoc adjustment.

The second big win is that {z'_i} can be negative even when {z_i > 0}—the swing term overpowers the other. This means the model projects the inferior move {m_i} as more likely than the engine’s optimal move. This is nervy but in many cases my model correctly “foresees” the player taking the bait.

The third big win—but tantalizing—is that the extended model not only allows solving 2 more equations but often makes other fitness tests {F'(s,c,h,a)} align like magic. The first of the following choices of extra equations makes an unbiased estimator for the frequency of playing a move of equal value to {m_1}, which became my third cheating test after its advocacy in this paper (see also reply in this):

  • {\mathsf{EV}_p = \frac{1}{T}\sum_{t=1}^T \sum_{i: v_{i,t} = v_{1,t}} p_{i,t}}.

  • {\mathsf{Error}(a,b)_p = \frac{1}{T} \sum_{t=1}^T \sum_{i: a \leq \delta(v_{1,t},v_{i,t}) \leq b} p_{i,t}}, for various error bounds {a \leq b}.

  • {\mathsf{M_jM}_p = \frac{1}{T}\sum_{t=1}^T p_{j,t}}; {j = 2} or {3} or {4\dots}

  • {\mathsf{IndexFit} = \sum_j w_j |\mathsf{M_jM}_p - \mathsf{M_jM}_a|^2} for appropriate weights {w_j}.

  • The frequency of certain kinds of move, e.g. moving a Knight, or capturing, or retreating.

A typical fit that looks great by all these measures is here. It has 26,450 positions from all 497 games at standard time controls with both players rated between Elo 2040 and Elo 2060 since 2014 that are collected in the encyclopedic ChessBase Big 2017 data disc. It shows {\mathsf{M_jM}} for {j=1} to {15}, then {\mathsf{IndexFit}} and tests related to it, then {\mathsf{MM}} is repeated between {\mathsf{ASD}} and {\mathsf{ETV}}, and finally come four cases of {\mathsf{Error}(a,b)} for 0.01–0.10, 0.11–0.30, 0.31–0.70, and 0.71–1.50, plus four with {b = +\infty}.

Only {\mathsf{MM}}, {\mathsf{M2M}}, {\mathsf{ASD}}, and {\mathsf{ETV}} were fitted on purpose. All the other tests follow closely like baby ducks in a row, except for some like captures and advancing versus retreating moves where human peculiarities may be identified. The {\mathsf{IndexFit}} value of {0.0052} is 5–10x as sharp as what my old model typically achieves. The new model seems to be confirming itself across the board and fulfilling the goal of giving accurate projected probabilities for all moves, not just the best move(s). What could possibly be amiss?

Whiplash Under the Dial

The first hint of trouble comes from the fitted value of {s} being {0.0083}. In my old model, players rated 2050 give {s} between {0.13} and {0.14}, while even the best players give {s > 0.06}. Players rated 2050 are in amateur ranks and {s < 0.01} leaves no headroom for masters and grandmasters. The {h} value of {1.80} compounds the sharpness; together with {s}, a slight value difference {\delta(v_1,v_i) = 0.05} (say) gets ballooned up to {10.84}, giving {z_i \approx 6.5^c = 2.50\dots} and {p_i = p_1^{e^{2.50}} = p_1^{12.20}}, which shrinks near 1-in-5,000 when {p = 1/2} and below 1-in-650,000 when {p_1 = 1/3}. This is weirdly small—and we have not even yet involved the effects of the swing term with {\rho_i}.

Those effects show up immediately in the file. I skip turns 1–8, so White’s 9th move is the first item. In the following position at left, Black has just captured a pawn and White has three ways to re-take, all of them reasonable moves according to the Stockfish 7 program.

Positions in game Franke-Doennebrink, 1974 at White’s 9th move (left) and Black’s 11th (right).

Here is how my new model projects them:

NRW Class1 1314;Germany;2014.02.02;6.4;Franke, Thomas;Doennebrink, Elmar;1-0
r1b1k2r/pp2bppp/2n1pn2/3q4/3p4/2PBBN2/PP3PPP/RN1Q1RK1 w kq - 0 9; c3xd4, engine c3xd4
Eval 0.24 at depth 21; swap index 1 and spec AA2050SF7w4sw10-19: (InvExp:1), Unit weights with
s = 0.0083, c = 0.3846, d = 12.5000, v = 0.0500, a = 0.9863, hm = 1.8024, hp = 1.0000, b = 1.0000:

M# Rk   Move  RwDelta  ScDelta  Swing  SwDDep  SwRel   Util.Share   ProjProb'y
 1  1  c3xd4:   0.00   0.000   0.000   0.000   0.000            1   0.79527569
 2  2 Nf3xd4:   0.42   0.321  -0.035  -0.034  -0.034     0.144422   0.20472428
 3  3 Be3xd4:   0.55   0.395   0.008   0.005   0.005   0.00445313   0.00000001

That’s right—it gives zero chance of a 2050-player taking with the Bishop, even though Stockfish rates that only a little worse than taking with the Knight. True, human players would say 9.Bxd4 is a stupid move because it lets Black gain the “Bishop pair” by exchanging his Knight for that Bishop. Of 155 games that ChessBase records as reaching this position, 151 saw White recapture by 9.cxd4, 4 by 9.Nxd4, and none by 9.Bxd4. So maybe the extremely low projection—for 9.Bxd4 and all other moves—has a point. But to give zero? The {0.00445313 \approx 1/225} is the {u_i} utility share, so the {p_i} is actually about {0.795^{225} = 4\times 10^{-23}}; the {0.00000001} is an imposed minimum. My original model—setting {h = 0} and fitting only {s} and {c}—spreads out the probability nicely, maybe even too much here:

 M# Rk   Move  RwDelta  ScDelta  Swing  SwDDep  SwRel   Util.Share   ProjProb'y
  1  1  c3xd4:   0.00   0.000   0.000   0.000   0.000            1   0.57620032
  2  2 Nf3xd4:   0.42   0.321  -0.035  -0.034  -0.034     0.280586   0.14018157
  3  3 Be3xd4:   0.55   0.395   0.008   0.005   0.005     0.241178   0.10168579

At Black’s 11th turn, however, the new model gives three clearly wrong “zero” projections:

NRW Class1 1314;Germany;2014.02.02;6.4;Franke, Thomas;Doennebrink, Elmar;1-0
r1b2rk1/pp2bppp/2nqpn2/8/3P4/P1NBBN2/1P3PPP/R2Q1RK1 b - - 0 11; Rf8-d8, engine b7-b6
Eval 0.11 at depth 20; swap index 1 and spec AA2050SF7w4sw10-19: (InvExp:1), Unit weights with
s = 0.0083, c = 0.3846, d = 12.5000, v = 0.0500, a = 0.9863, hm = 1.8024, hp = 1.0000, b = 1.0000:

M# Rk   Move  RwDelta  ScDelta  Swing  SwDDep  SwRel   Util.Share   ProjProb'y
 1  1  b7-b6:  -0.00  -0.000   0.000   0.000   0.000            1   0.56792559
 2  2 Nf6-g4:   0.18   0.163  -0.001  -0.002  -0.002    0.0907468   0.00196053
 3  3 Rf8-d8:   0.18   0.163   0.042   0.046   0.046   0.00391154   0.00000001
 4  4 Bc8-d7:   0.21   0.187  -0.029  -0.030  -0.030     0.278053   0.13071447
 5  5 Nf6-d5:   0.28   0.241   0.047   0.050   0.050   0.00218845   0.00000001
 6  6  a7-a6:   0.30   0.256  -0.049  -0.051  -0.051      0.28777   0.14001152
 7  7 Qd6-c7:   0.31   0.264  -0.012  -0.012  -0.012     0.097661   0.00304836
 8  8  g7-g6:   0.37   0.306   0.015   0.017   0.017   0.00355675   0.00000001
 9  9 Qd6-d8:   0.39   0.320  -0.054  -0.051  -0.051     0.206264   0.06438231
10 10 Qd6-b8:   0.39   0.320  -0.037  -0.038  -0.038     0.158031   0.02787298

Owing to many other games having “transposed” here by a different initial sequence of moves, Big 2017 shows 911 games reaching this point. In 683 of them, Black played the computer’s recommended 11…b6. None played the second-listed move 11…Ng4, which reflects well on the model’s giving it a tiny {p_2}. But the third-listed move 11…Rd8 gets a zero despite having been chosen by 94 players. Then 91 played the sixth-listed 11…a6, which actually gets the second-highest nod from the model, and 22 played 11…Bd7, which the new model considers third most likely. But 12 players chose 11…Nd5, four of them rated over 2300 including the former world championship candidate Alexey Dreev in a game he won at the 2009 Aeroflot Open. My old model’s fit of the same data gives 34.8% to 11…b6, 10.4% to 11…Ng4 and 7.5% to 11…Rd8 with the ad-hoc change for tied moves (would be 8.7% to both without it), and 5.1% to 11…Nd5, with eighteen moves getting at least 1%.

To be sure, this is a well-known “book” position. The 75% preference for 11…b6 doubtless reflects players’ knowledge of past games and even the fact that Stockfish and other programs consider it best. It is hard to do a true distributional benchmark of my model in selected positions because the ones with enough games are exactly the ones in “book.” Studies of common endgame positions have been tried then and now, but with the issue that the programs’ immediate complete resolution of these endgames seems to wash out much of the progression in thinking and differentiation of player skill that one would like to capture. (My cheating tests exclude all “book-by-2300+” positions and all with one side ahead more than 3.00.) Most to the point, the fitting done by my model on training data is supposed to be already the distributional test of how players of that rating class have played over many thousands of instances.

The following position is far from book and typifies the most egregious kind of mis-projection:

SVK-chT1E 1314;Slovakia;2014.03.23;11.6;Debnar, Jan;Milcova, Zuzana;1-0
2r4k/pp5p/2n5/2P1p2q/2R1Qp1r/P2P1P2/1P3KP1/4RB2 b - - 1 32; Qh5-g5, engine Qh5-g5
Eval 0.01 at depth 21; swap index 2 and spec AA2050SF7w4sw10-19: (InvExp:1), Unit weights with
s = 0.0083, c = 0.3846, d = 12.5000, v = 0.0500, a = 0.9863, hm = 1.8024, hp = 1.0000, b = 1.0000:

 M# Rk   Move  RwDelta  ScDelta  Swing  SwDDep  SwRel   Util.Share   ProjProb'y
  1  1 Qh5-g5:   0.00   0.000   0.129   0.137   0.000    0.0347142   0.00018527
  2  2 Rc8-d8:   0.00   0.000   0.026   0.025  -0.112            1   0.74206054
  3  3  a7-a6:   0.09   0.085  -0.008  -0.005  -0.142     0.117659   0.07922164
  4  4 Rc8-g8:   0.21   0.187  -0.092  -0.087  -0.225    0.0996097   0.05003989
  5  5 Rh4-h1:   0.25   0.219  -0.166  -0.165  -0.302     0.136659   0.11270482

This has two tied-optimal moves for Black in a position judged +0.01 to White, not a flat 0.00 draw value, yet the one that was played gets under a 1-in-5,000 projection. Here are the by-depth values that produced the high positive {\rho_1} value:

         5    6    7    8    9   10   11   12   13   14   15   16   17   18   19   20   21
Qg5    -117 -002 -008 +000 -015 +008 +017 +032 +011 +007 +000 +004 +001 +000 +000 +006 +001
Rd8    -089 -058 -036 -032 -013 -025 +006 +000 -010 +000 -012 -013 +001 +014 +001 +001 +001

The numbers are from White’s view, so what happened is that 32…Rd8 looked like giving Black the advantage at depths 10, 13, and 15–16, whereas 32…Qg5 looked significantly inferior (to Stockfish 7) at depth 12 and nosed in front only at depth 20 just before falling into the tie. The swing computation begins at depth 10 to evade the Stockfish-specific strangeness I noted here last year, so in particular the “rogue” values at depth 5 (and below) are immaterial. The values and differences from depth 10 onward are all relatively gentle. Hence their amounting to a tiny {u_1} versus {u_2} and microscopic {p_1} is a sudden whiplash.

A Kind of Race Condition

What I believe is happening to the fit is hinted by this last example giving the highest probability to the 2nd-listed move. Our first game above has two positions where the 9th-listed move gets the love. (The second, shown in full here, is notable in that the second-best move gets a zero though it is inferior by only 0.03 and was played by all three 2200+ players in the book.) This conforms to the goal of projecting when weaker players will prefer weaker moves.

This table shows that the new model quite often prefers moves other than {m_1}, compared to how often they are played:

\displaystyle  \begin{array}{|r|r|r|} \hline \text{Index} & \text{Preferred} & \text{Played}\\ \hline 1 & 14,808 & 12,388\\ 2 & 5,874 & 4,738\\ 3 & 2,335 & 2,506\\ 4 & 1,111 & 1,556\\ 5 & 725 & 1,117\\ 6 & 423 & 752\\ 7 & 317 & 593\\ 8 & 226 & 501\\ 9 & 145 & 368\\ 10 & 126 & 334\\ \hline \end{array}

To be sure, the model is not putting 100% probability on these preferred moves, but when preferred they get a lot more probability than under my old model, which never prefers a move other than {m_1}. Recall however that my old model’s fit was not too far off on these indices—and both models are fitted to give the same total probability to {m_{1,t}} over all positions {t}. Hence the probability on inferior moves is conserved but more concentrated.

Yes, greater concentration was the goal—so as to distinguish the most plausible inferior moves. But the above examples show a runaway process. The new model seems to be seizing onto properties of the {\rho_i} distribution alone. For each {t} we can define {\mu_t} to be the move {m_{i,t}} with the most negative value of {\rho_i - \rho_1}. The {\mu_t} also form a histogram over {t}. The fitting process can grab it by putting all weight on {\mu_t} plus at most a few other moves at each turn {t}.

These few moves are the “stopped-watch reading” in my analogy. The moves given zero are the readings that cannot happen for a given runner/position. The fitting doesn’t care whether moves getting zero were played, so long as other turns fill in the histogram. If a high {p_2} for {m_2}—as with 32…Rd8 above—fills a gap, the fit will gravitate toward values of {s} and {h} that beat down all the moves {m_{i}} with {i \neq 2} at such turns {t}. In trials on other data, I’ve seen {s} crash under {0.0001} while {h} zooms aloft in a crazy race.

What can fix this? The maximum likelihood estimator (MLE) in this case involves minimizing the log-sum {\sum_t \log(1/p_{i_t,t})} of the projected probabilities of the moves {m_{i_t}} played at turn {t}. Adding it as a weighted component of {F(s,c,h,a,\dots)} helps a little by inflating the probability of the moves that were played, but so far not a lot. Even more on-point may be maximum entropy (ME) estimation, which in this case means minimizing

\displaystyle  \frac{1}{T}\sum_{t=1}^T \sum_i p_{i,t} \log(p_{i,t}).

There are various other ways to fit the model, including a quantiling idea I devised in my AAAI 2011 paper with Guy Haworth. In principle, and because the training data is copious, it is good to have these ways agree more than they do at present. Absent a lightning bolt that fuses them, I am finding myself locally tweaking the model in directions that optimize some “meta-fitness” function composed from all these tests.

Open Problems

Is this a known issue? Does it have a name? Is there a standard recipe for fixing it?

Do any deployed models have similar tendencies that aren’t noticed because there isn’t the facility for probing deeper into the grain that my chess model enjoys?

[added “at standard time controls”, a few other word changes, added game diagrams]

One Comment leave one →
  1. Ioannis Avramopoulos permalink
    May 29, 2017 6:22 am

    Very nice post.

    Definition: Consider a 2-player game, say G, of alternating moves/rounds (e.g. chess) with three possible outcomes, namely, win/draw/lose (like chess). Define G to be R-complete if there exists a Turing machine in the position of a player scoring at least draw (or better) provided the Turing machine is allowed at most a polynomial number of steps in each round.

    I think Zermelo showed that chess is always a draw for ‘infinite’ players.

    Conjecture: Chess is R-complete.

    *R in R-completeness is for Reagan.

    **I am writing from Dodoni coffee shop in Melissia, Athens.

    ***No regret algorithms in machine learning provide guarantees against arbitrary intelligence. Ken’s post made me think that this is quite important: No matter how infinitely human you are, there are Turing machines that cannot beat you in some environments.

Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s