Thursday, February 25, 2016

Local Maxima

It’s weird when I feel those bursts of inspiration to improve at Melee again. Also weirder when I write about them in a blog.

I just came back to America from my first international tournament, BEAST6, in Sweden. It was a wonderful time, and I’ve had the chance to reflect on a lot of things. One of those things is that I got 7th place, and I would much rather have won. Another is where I’m trying to take my training and development in the game.

When I tell people that I am trying to win and get better, their first response is usually “start wobbling again.” Normally, my response has been “but it’s boring” or “I don’t like playing like that anymore.” I’m starting to take something of a different attitude right now; this newer attitude is based on looking back at my thoughts and emotions while playing, as well as the fruits of my different playstyle.

*

You are blindfolded, standing on a hill. You can feel wind whipping around your face, and you gingerly put your foot out, trying to find ground to walk on. After spinning around a bit, you find a direction that is sloping upward, and you start to climb the hill.

It’s slow going, but you poke around and feel with your hands and eventually it feels like you have climbed a fair distance. Finally, you reach a point where nothing leads up, and you are at the peak of the hill. Happy, you take off your blindfold and look around to survey your progress.

It turns out that you are actually at the base of a massive mountain, and its peaks are soaring high into the clouds. Your hill, by comparison, is incredibly tiny.

Your hill can be called a “local maximum.” It’s the highest point in its immediate vicinity. Every change from your current location would be a downgrade. But if you ever want to climb the mountain, you will have to leave the hill, and you will have to go downward. If you want to go up, you have to go down.

This is, funnily enough, a problem in machine learning. A program will learn how to do something, and improve its results, and then get stuck doing something “optimal.” Because any immediate change will reduce its results, it never changes, repeating itself. However, it only found an “okay” strategy or idea and perfected it; it has to go in a much different direction to achieve truly great results. When you want a computer to do the learning for you, you need algorithms that don’t get it caught up in local maxima.

(Recent book: The Master Algorithm by Pedro Domingos. It’s really good!)

Though there are very bright people who have found ways to handle that problem, many non-machine learners get stuck in it as well. Currently, this is the problem I’ve been thinking about in Melee with regards to my character.

Right now, most ICs will use the infinite to secure kills. They will also use styles that emphasize the grab game, to get as many of those infinites as possible. This works pretty well against a lot of people, and consequently, Ice Climbers that wobble a lot will win a lot.

Unfortunately, it seems like we’ve hit a peak. When somebody is skilled at avoiding that grab… what do we do? Well, generally speaking, we lose. Our answer is typically “we have to get better at grabbing somehow,” and despite that we are not getting that much better. In fact, certain players absolutely eviscerate us, leaving our sad eskimo entrails spilling all over loser’s bracket.

We have pretty much reached the peak of that hill, but the real mountain is sitting right next to us, tall and imposing as ever before. It also makes people think that the character lacks the potential to succeed. You just have to look at how badly the best players destroy us.

The real answer, I think, is to start climbing down, so we can start climbing up.

*

Originally, I just wanted to avoid wobbling so I could do crazier, sillier things with my character and have more fun. Now I am actively trying to avoid thinking about it because I want to attain the skills that other players have, and apply them to my character.

One problem is that our infinite is just *too rewarding.* It skews your thoughts towards getting that next grab. Even if it doesn’t, it can make you complacent, letting you ignore certain mistakes because you’re still winning by a fair amount. It also causes you to avoid certain difficult situations, because they won’t lead to grabs, so why bother? Especially when those positions are risky!

However, this means we are ignoring facets of the game that every other player has to master. We aren’t focusing on our defensive smash DI. We aren’t focusing on the close-quarters scuffling that every Fox learns because he has a Falco for a training partner. We make lots of little technical mistakes, or simply fail to optimize them, because those don’t necessarily pertain to our grab game, and then they cost us.

In short, we are falling behind. We are falling below. Other people are climbing up the other peaks, and we are idly shuffling around, trying to find the best spot on our hill to watch them.

Another problem our character has is that… well, our character is very very different from the others. We have a different neutral, one that is heavily based on the wavedash, like Luigi, rather than dashes or air mobility. ICs are also the only puppet character, which involves desyncs, understanding the other Climber’s AI, and playing miniature 2v1s constantly during matches. We have a group of skills to master that no other character worries about, which spreads our attention. Lest I sound like I'm ragging on other ICs too much, that isn't our fault at all, and it's part of why we're struggling.

Again, this means that while other players are using smash DI on jab-resets to avoid guaranteed followups, or practicing reaction tech-chases, or mastering invincible wavelands, we are still trying to figure out the mystery of how Nana knows the exact worst time to taunt every match.

We are playing a different game. Our emphasis on grabs (with a character that itself is bad at grabbing skilled players) is making it a much smaller game. The infinite also tricks us by giving that small game a huge payout, making us think the hill is the peak of the mountain. But it’s not.

*

So enough about us, let’s talk about me.

One of my current weaknesses is that I’m trying to play extremely fast all the time. I am currently trying to push my technical skill and eke out every frame from every situation I can. I am trying to use invincible wavelands (and SD’ing). I am trying to master pivots for the small benefit they give in close quarters microspacing. I am also trying to add this to quicker reactions, particularly in scuffles (those situations where two players are on each other in close range, hitting lots of buttons).

The result is that I’m making a lot more errors. These errors are costing me games and sets, and I’m just going to have to deal with that. You have to go down the smaller hill to start climbing the bigger mountain. Other players are using these kinds of advanced tools and using them consistently, so I have no excuse for not integrating them.

I am also trying to avoid wobbling as much as possible. If I get a grab, I intentionally convert it into a different scenario. Originally this was just for funsies, but now I think that it is, in fact, the right thing to do if I want to improve. Mind you, there is no point in truly forsaking such a strong punish game. It kills people dead. As far as punishment strategies go, it is pretty much the peak for my character, whenever it’s available. That is why this past weekend, I tried to use it as often as I could, especially in my (super close) top 8 set against Professor Pro. I am still highly competitive. I would like to win!

However, if I use it every time, particularly in tournaments and pressure situations, I’m depriving myself of the opportunity to train other skills in those situations. If I keep doing that, the result will look a bit like this:

“Okay, tourney time, I’m going to use the stuff that’s consistent and not give away any openings.”
“Oh shoot, I just flubbed that pivot.”
“Uh oh, my smash DI and CCs are pretty bad, I should probably play on the outside more.”
“I can’t seem to get these autocancel u-airs right now, I’ll stay grounded.”
“Well, at least my wobbling still works, I’ll just focus on that.”

I will just end up doing what I'm most comfortable with, which is my old stuff. And it will keep happening, because if I don't train the new stuff under pressure, I will botch it under pressure. I will only incentivize myself to keep doing the same old thing, the same old strategy that will not keep working as my opponents improve. As they get better, I will become more and more paranoid about taking a few steps down my little hill.

Somewhere in here is a happy medium for pushing myself without sandbagging pointlessly and developing lazy habits with regards to my grab game. I'll let you know when I find it.

Finally, this is a list of some things I’m trying to work on. You will notice that this list is actually kind of long. Most of the stuff is also hard.

--Invincible ledgedashes
--Smash DI’ing small hits into counter hits, or to escape
--Shield angling to avoid pokes
--Shield impact DI to guarantee punishes or evade followups
--Tighter on-shield actions to stay safe, like aerial->jab
--Tighter control over aerial drift, for staying safer, dodging hits, and placing aerials better
--Better reactions in tech-chase scenarios
--Implementing effective desyncs in neutral
--Implementing faster desyncs in neutral so I don’t lose my setups under enemy pressure; also using hit-stun and shield-stun desyncs on purpose to generate advantages out of nowhere.
--Smarter CC’ing--making it intentional so I don’t DI wrong, but implementing it intelligently to score counter hits and earn respect for my space
--Tighter dash dance game, especially around get-up attacks and landing aerials.
--Intentionally clinking with moves to force my way into range or prevent enemy dominance over the footsies game
--Accounting for enemy hitstop on Nana and avoiding those extended hitboxes when counterpoking
--Tracking Nana’s AI in fast-paced situations and remembering her various methods of getting me into trouble
--Avoiding extended shield-based play, since it’s mostly just bait for shield-grabs, but it increases the odds of getting shield-poked.
--Not getting into crummy situations in the first place so I don’t need to use risky tech-skill to bail myself out constantly

I could probably think of a few others but those are the ones that came to mind off the top of my head. This is aside from simply trying to up the consistency on other skills I already have. If it seems ridiculous, that’s probably just because our game is ridiculous. More importantly, the best players also use all of those skills (well, the non-IC-specific ones) and use them consistently because that’s what gives them an edge. If I want to step into their realm and keep ascending with them, then I need to gain access to those skills as well.

The downside is that it means willfully entering situations when using tools I lack experience with. This means more errors, more free losses, and sometimes throwing away things that I know will work in favor of stuff that I hope will pay off. It means losing matches I might have won if I’d played safer, more within my comfort zone. It also means more frustration as I commit unforced errors, but that’s just another weight to carry up the mountain with me.

Those are some of my thoughts lately. I am very determined to start training a lot more and try to achieve some of these in my game, and hopefully show people we’ve still got a long way to climb. Thanks for reading.