Posted on 03 May 2014 - 12:53 AM

I have a few observations and suggestions regarding the ranking system.

[size=200]Inconsistency and Skill[/size]

Some ranked ladders have luck issues; that cannot be denied. Vanilla's Strength II critical hit mechanics and the nature of Unbreaking III can turn the tide of a fight when someone deals 3.5 hearts of damage with a single hit (minimal damage reduction & maximum critical damage), or when their boots break before their opponents despite their damage output being greater than that received.

The ranking system's Elo K-factor (maximal Elo difference between pre- & post-fight ranking) is currently set around 30. A 2000 Elo aficionado loses to a series of lucky critical hits from a 1500 on a roll? They lose almost 30 points, and that could take them more than 15 games to replenish. It's a pretty harsh punishment for what can come down to a series of unfavourably generated numbers.

BadLion's ladders heavily punish inconsistency because of the high K-factor. The Hundred Elo Drop isn't nearly as fun as the Hundred Acre Wood – lose two games, and you've lost almost 50 Elo. Continuous matching with people who are labelled with a lower skill level is going to result in occasional unexpected results, and that's fine: the system will eventually converge on an accurate value for a participant, but as random numbers get more involved and unexpected results occur more often, a high K-factor creates inaccuracies at the higher end of the spectrum… the part that we actually look at!

[size=200]Matchmaking[/size]

BadLion doesn't have this. BadLion should have this. When matches are often heavily skill-imbalanced, I don't think that the stompers enjoy it, and similarly, I don't think that those being stomped enjoy it much either. Not having matchmaking of any sort greatly increases the likelihood of flukey results.

I'll gloss over a couple of methods of matchmaking that I can think of:

Queue Partitioning

N people queue for fights every T seconds. The queue is sorted and partitioned into pairs, which represent the matches, so that you'll always be put against someone with a reasonably close Elo to your own, although the quality of the match made depends on queue's population. This is a simpler option that scales reasonably regardless of queue population.

Bracket Queues

Multiple queues are set up based on rank intervals with hard boundaries. This method would increase queue times for people at the highest and lowest ends of the spectrum, but in contrast to queue partitioning, a high skill discrepancy match (e.g. 2000 vs. 1500) can never be made. Results of this method would be the most accurate.

[size=200]Solutions[/size]

1. Decrease the K-factor for the vanilla queue (at least). I'd recommend halving it to ~15.
2. Implement matchmaking and tune the K-factor accordingly. A higher K-factor works when the rank disparity of matches is smaller.
3. Modify the mechanics of the queues with the most random factors. Or don't. I know you won't.

Posted on 03 May 2014 - 12:53 AM

y u no size tag

Posted on 03 May 2014 - 01:01 AM

not gonna quote this whole thing, but +1 to basically everything lol

Posted on 03 May 2014 - 01:01 AM

Posted on 06 May 2014 - 02:38 AM

If the server was massive enough, I think match-making could work. I don't see it working, especially at Australian timezones, for now.

+1 to everything else