How Machine Learning and AI are Changing eSports and Knowledge Itself
The crowd at Key Arena was roaring as one of the greatest stars in eSports, Danil “Dendi” Ishutin, took the stage. His then-team, Natus Vincere–– affectionately known as Na’Vi––hadn’t qualified for The International 7, or 2017’s largest Dota tournament. Despite the disappointing result in the qualifiers, Dendi had once again found himself on the main stage of The International where he had competed in an unheard of three consecutive grand finals, winning one title in 2011. Everyone was cheering for Dendi in what we would soon discover to be a showmatch between himself and his to-be-announced opponent, likely another legendary player. As he entered, he lowered his hood and explained with his customary cheer that he was ready for the challenge.
Dendi has this absolute sense of star power, an ineffable, indefatigable charisma that will charm the hell out of anyone even if they have no idea what “that weird computer game you play all the time” is, according to a friend. That charisma is also matched by some incredible skill at a really challenging game. Dota is notorious for taking years to even learn the basics of. At almost five thousand hours, I’m just starting to understand it myself. It’s for this particular reason that what happened next was so incredible.
As Alex “Machine” Richardson, co-host at The International hyped Dendi and the crowd up, he then announced his opponent was ready. The crowd, however, was not. In came a full team, clad in shirts emblazoned Security. In the center, they wheeled with them a desktop tower. After they had plugged in the tower, they dramatically plugged in a USB stick. Game, set, match. The showmatch was ready.
This game of Dota was just one versus one, a contest of pure skill. Dendi, traditionally a high-skilled playmaker from the middle lane––often the “purest skill” lane of the three in Dota––was matched up against a prototypical artificial intelligence, OpenAI. A company co-founded by Elon Musk, which he later left but still supports, OpenAI describes itself as “a non-profit AI research company, discovering and enacting the path to safe
artificial general intelligence.” The company began training bots in Dota 2, clearly looking to outperform Google DeepMind’s AlphaGo project. And outperform it did.
After one hard-fought loss, Dendi entered the second game trying to outplay the AI bot––ultimately finding it to be better than him. “Okay, I give up.” he said, conceding the match before it was even over. For me, sitting in Key Arena, it dawned on me that Dota would now be played differently. AI, specifically OpenAI’s work in that moment, would be a tool for improvement. It could do things that humans simply couldn’t do, at least not at any reasonable rate of progress.
The bot had trained meticulously over thousands upon thousands of games of Dota, at first just trying to master the mechanics of the game, but eventually mastering mechanics that even veterans of the game had only just begun to scratch the surface of. These bots trained with themselves at a rate of 180 years of gameplay per day. The average high-skilled Dota player, in order to stay competitive, gets in anywhere from six to twelve games per day, allowing for the need to sleep. OpenAI possessed a superior practice regimen, allowing it to far exceed what humans could ever hope to do.
And surpass humans it continued to do––as in August of 2018, when the OpenAI Five, a complete five-man, regular Dota roster of bots faced a set of high-skilled humans. Earlier in June, the team had released a proof of concept video, winning applause from Bill Gates. The bots in August won, with one member of the human team, a high-level analyst named William “Blitz” Lee explaining, “The teamwork aspect of the bot was just overwhelming. It feels like five selfless players that know a good general strategy.” These bots had fundamentally changed the way Dota was understood. There was no going back: Dota would now be played on different terms.
The OpenAI team’s bots made moves that left most conventionally high-skilled Dota players––myself included––totally baffled. Dropping a sentry ward to draw tower aggro so you didn’t take any hits while forcing an objective on the map? Next level play. I was amazed by the amount of skill these bots showed, but then again, they had years upon years of random input-based experience to push them ahead of the competition. It’s for this reason in particular that AI will become such a strong analytical tool for eSports pro-gamers. It changed Dota, and I’m confident that it will continue to change the way games themselves are played.
The Nature of Competition
If we look at the difference between any high-level athlete and a basic athlete, there are a few dimensions: efficiency, strategy, and, in the case of a team sport, team camaraderie. Efficiency is the most glaringly obvious piece––great Olympic rowers have efficient, lean strokes which utilize the perfect amount of strength based on their form, while Dota players have an efficient mindset when approaching the game. They are one and the same, rowing just happens to be more physically taxing. Strategy is also something high-level competitors enter any match, race, or game with: they have a clearly-defined plan for victory. For rowers, it’s when to drive at the precise moment to capitalize on an advantage. For Dota players, it’s knowing how to draft and execute on that draft perfectly to defeat an opponent. Finally, team camaraderie is the least salient piece to any sport because it’s just assumed: good teams must be on good terms with one another. The best rowers are in-sync with one another and tend to enjoy each other. The same goes for Dota players––as William “Blitz” Lee said, there’s a degree of selflessness to the game that one needs to play with.
AI helps with two of those areas, and perhaps can help with the third in the future. But for now, it serves as a huge boon to progamers’ efficiency and strategy. The move the OpenAI team pioneered which I had mentioned earlier was an example of extreme efficiency. Lo and behold, top-tier players began to pick up on it. The bots’ strategies were also unconventional, which will inevitably lead players to consider the game in a different light when it comes to viable ways to play; this is especially true in a game replete with as much complexity as Dota.
As further evidence of this transformation, the winners of The International 2017, Team Liquid, partnered with German company SAP SE (Systems, Applications, and Products in Data Processing) to create a data analytics tool to help them better understand the game. Unsurprisingly to many, Liquid went on to dominate the rest of the tournament scene, ultimately placing fourth at The International 2018 after a series of strong placements and a major victory. SAP and Team Liquid have continued their partnership, and it will be supremely interesting to see how the team fares in the upcoming major in a few weeks.
The Limits of Human Knowledge
With machine learning for data analysis and AI to test new strategies in the game, the field for competition is widening in eSports. OpenAI and its team, along with SAP’s data analytics tools using SAP Leonardo, have begun to push the limits of a game most thought to be understood. What AI writ large reveals to us, then, is a fundamentally human understanding of reality: patterns are followed and iterated upon, but there is still a singular narrative which is followed. What the OpenAI team has taught us in Dota is that though the game is played much the same way, there are a multiplicity of other, more varied narratives to explore.
It’s one of the more positive applications of philosopher Jean Baudrillard’s argument that the simulation of the real will begin, increasingly, to precede the real. Our knowledge base being so limited by human constraints of temporality, mortality, and bias, machines––the simulators––offer several alternative views of how situations can be approached. We should, then, embrace some aspects of simulation. Other parts can go, for sure, but expanding human knowledge seems to be a net positive effort.
We may never be able to compete with a machine that can process human experiences at efficiencies upward of 180 times more than ourselves. However, we can learn from those machines which provide us with such rich data. Just as Dota players are learning daily from the new information OpenAI or SAP bring to the table, as AI itself advances, there will be dramatically new understandings of the narratives of reality which we will be able to explore––understandings of our world we have been totally blind to due to our own biases based on an iterative but nevertheless comparably small knowledge base. It’s that prospect that is so exciting: that much like in Dota, there is so much more room to grow––that there are so many stones left unturned.