Dota 2: What lessons can we learn from OpenAI?

Many members of the Dota 2 community misunderstand the purpose of OpenAI – they think it was made to play Dota. But that’s not enough. Dota is a place for developers to test their intelligence – to see if their algorithms can solve problems. And the results they obtained are quite positive.

The interesting thing about this testing process is that the AI ​​uses almost the same tactics as humans have been doing for the past 15 years. One mid hero, two or three lane heroes, rune control, etc – it’s similar to what we humans do when we play Dota. It played out like a normal game.

Of course, OpenAI can coordinate, centralize opponents as well as make decisions to a degree that is impossible for humans and that is not the focus of this article. There are some things AI does differently from us. Those are the things that we players need to consider as well as review the way we play Dota ourselves.

GO LANE EARLY GAME

Players often prefer to use stability during the laning phase. You choose a lane and then I hope to meet a weaker opponent. Maybe, you will change lanes once and then mostly farm there, trade hits and do the usual things in Dota. OpenAI doesn’t endorse that way of playing.

Looking back at the replays of both games, it’s easy to see that the AI ​​changes lanes very often: rotates in and out constantly to find an advantage for the team. As soon as some heroes reach level 3, they start looking for opportunities to pressure and punish their opponents. AI doesn’t matter who the hero rotates: core or support – what matters is the success of the rotation.

Theoretically it’s easy to understand: if we rotate a hero out of dual lane, you still have one hero in lane to gain experience, so you don’t really lose. Killing enemy heroes is almost always more profitable than the cost of a TP. Then after ganking, you can rotate a hero or two, maybe find another gank on the other side of the map.

Of course, humans are limited in this range. We cannot accurately calculate the success or expected value of rotate with no rotate. But we’ve seen similar gameplay since 2015, coming from a team called CDEC and their star Agressif. Coincidentally, his hero in the tournament was also Gyrocopter and Agressif kept rotating, optimizing his effectiveness. His playing style always brings success: no one can say that the TI runner-up position was due to luck.

FARM . DISTRIBUTION

The results of the chaotic laning phase show that the OpenAI team’s level and gold are more balanced. But the most important thing is that OpenAI’s overall Net worth is higher than the player’s team (OG).

Unfortunately, the observers don’t show us the experience advantage of the first two teams, but only by level, with AI always having the advantage. OpenAI has no “sacrifice” support, runs around the map for 10 minutes with only level 3 and can’t get himself boots (or more accurately, spend money to buy wards). OpenAI prefers 5 heroes who are capable of fighting and ready for any situation.

Dota has a concept of “playing around a hero”: professional teams in the current meta often let the two controlling heroes go live early and create plays with support from around the map. OpenAI simply turns every hero they have into a “playable” hero, at least until the first 15-20 minutes of the game. This greatly increases the success of OpenAI.

Obviously, over time, the imbalance starts to become more pronounced and the heroes who farm better get more resources, but this imbalance is not as pronounced as on the human side. AI doesn’t want to ask “top net worth” heroes to help them win matches. They just wanted to maximize the probability of winning and decided resource allocation was the best way to do this.

BUYBACK AND EXPERIENCE PRIORITY

Another interesting point that we need to mention is that OpenAI is very fond of XP. For that reason, at a certain point, they start buying back regularly and don’t hesitate at all. Buybacks, TP to shrines and jungling are signals that a player is often ’tilt’, but the AI ​​isn’t toxic or anything. It just wants to win.

Starting around level 9, the AI ​​will buyback instantly on almost any hero to work around the map. The balanced net worth ensures that buybacks from the “core” are not too expensive, while the OpenAI lineup always has a way to push waves or farm quickly.

This helps the AI ​​always keep the XP advantage, almost always push the lane and fight 5 people at all times.

EPILOGUE

We can’t deny OpenAI’s victory, mainly due to its ability to make accurate decisions lightning fast. Look at how they focus on attacking objects, how to cast spells and calculate how long they can last in teamfight: it’s really complicated to think what we humans can learn from them.

Non-AI players: we don’t know exactly how much HP we have left after the enemy casts an ability, or how much damage we will take while being disabled by a certain hero, etc.

The question is, given our knowledge as well as our imperfections, should humans try and apply the tactics that OpenAI used in Dota? AI knows its capabilities well through millions of trials, and we don’t know the exact cause or the ultimate reward. But from what we learned from the match between OpenAI and OG, players should experiment with how they play.

According to Dotabuff