IOI 2010 Languages — More than 102 points?

#	User	Rating
1	ecnerwala	3649
2	Benq	3581
3	orzdevinwang	3570
4	Geothermal	3569
4	cnnfls_csy	3569
6	tourist	3565
7	maroonrk	3531
8	Radewoosh	3521
9	Um_nik	3482
10	jiangly	3468

#	User	Contrib.
1	maomao90	174
2	awoo	164
3	adamant	162
4	TheScrasse	159
5	nor	158
6	maroonrk	156
7	-is-this-fft-	151
8	SecondThread	147
9	orz	146
10	pajenegod	145

I think IOI 2010 Languages is a very interesting (but also slightly annoying) problem — not only is it one of the few IOI problems that use machine learning, but the maximum score is 110 points instead of the usual 100.

Because of how unusual the problem is, there are relatively few people who try to solve it and far fewer with more than 100 points.

I've recently been trying to push the limits on how many points I can score. So far, I've managed to score 102 points on Yandex — the highest score I'm aware of.

My code for 102 points

#include <bits/stdc++.h>
#define FOR(i, x, y) for (int i = x; i < y; i++)
typedef long long ll;

#include "grader.h"

const int SZ = 100, LANGS = 56, MOD = 539653;

ll si[SZ], bi[SZ], tr[SZ], qu[SZ];
double cnt[LANGS];
int freq[LANGS][MOD];

inline double hyperb(int x) { return x / (x + 1.1); }

void excerpt(int *E) {
    FOR(i, 0, SZ - 3) {
        si[i] = E[i];
        bi[i] = ((si[i] << 16) + E[i + 1]) % MOD;
        tr[i] = ((bi[i] << 16) + E[i + 2]) % MOD;
        qu[i] = ((tr[i] << 16) + E[i + 3]) % MOD;
    }

    int best = 0;
    double best_sim = 0;
    FOR(i, 0, LANGS) {
        double sim = 0;
        FOR(j, 0, SZ - 3) {
            sim += hyperb(freq[i][qu[j]]) * 6;
            sim += hyperb(freq[i][tr[j]]);
            sim += hyperb(freq[i][bi[j]]) * 4;
            sim += hyperb(freq[i][si[j]]);
        }
        sim /= log(cnt[i] + 2);

        if (sim > best_sim) best = i, best_sim = sim;
    }

    int ans = language(best);
    cnt[ans]++;
    FOR(i, 0, SZ - 3) {
        freq[ans][qu[i]]++;
        freq[ans][tr[i]]++;
        freq[ans][bi[i]]++;
        freq[ans][si[i]]++;
    }
}

I found the correct weights by trial and error, so there's obviously a lot of room for improvement.

Here's my question: how high can we (realistically) go? A score of 103 points requires at least 93.6% accuracy, but my code only has an accuracy of 92.8%

dolphingarlic's blog