Machine Learning for Optimal Matchmaking

• Converts concerns to objective function (ex. ping, skill difference, time it took to make matches.)
• Create a model of how many people per second appear in given skill/ping blocks. This model is then used to know when to hold a player longer for a better match or they have the best they are likely to get.
• Mapping skill disparity as a percentile means everyone has a relatively smooth curve of opponents.
• Something about mapping predicted skill scores to percentiles on a normal distribution, multiplying by a scale and then only matching those within a gap. Supposed to keep the wait times constant between high/low/mid grade players.

TODO better examination of using percentiles than elos; its a curve used so people on the fringes fight other fringers but people on the bigger populations fight others in the larger pool.