Conclusion
Obviously, creating a perfect or a near-perfect matchmaking algorithm for millions of players would not be simple. Additionally, one that ensures a win rate of approximately 50% for everyone almost all the time throughout a season or a split would be more than challenging in all aspects.
A few of them would be the theory and the actual application and maintenance of such a system. The algorithm used would not only have to be theoretically sound and proven but also be kept updated as time passes as certain circumstances change, both in-game and outside the game. The right human resources would have to be allocated with the right equipment to constantly monitor all the processes for abnormalities and outliers. Despite these challenges, the work in this document has shown that the current matchmaking system is attempting to somewhat mimic the workings of the Perfect Family matchmaking.
However, even different patterns that look random may not be random or ‘random-ish in the same manner’. There may exist subtle differences and variations in certain values that make them ‘not the same type’ of random. The purpose of this document was to find out the extent to which these differences could be exploited.
After all the experiments, it is clear that a group of records may tell us which matchmaking algorithm was used. This is made easier if said group was influenced by one same policy. Moreover, if the group was influenced by a composite scheme of two or more matchmaking algorithms, it may still exhibit different traits when compared to values from collective statistical tests on the Perfect Family group. In short, Live Data may as well be somewhere in between the Perfect Family algorithms and StreakMM.
Unfortunately, there is a limitation. Determining what was used on an individual record basis is laden with more obstacles than when doing the same in groups. The biggest one is the fact that the small yet significant differences are shared across millions of records. This hinders any model trying to classify the maker of such patterns as there may be identical or extremely similar samples with different labels to them. Such omnipresence of certain records would often greatly slow down the learning process of a model being trained as it confuses them. This would be the main reason why precision or sensitivity scores for Live Data and PerfectFam are not as high as one would hope.
Next: FIN
Previous: Chapter 9