A historic event for both DotA 2 and eSports as a whole just took place. Developers working on the Open AI project just showcased their bot’s teamplay mechanics in a five versus five setting. The bots faced up against audience members at first, making light work of their opponents. Following that match, a best of three series was played against prominent community members. “Team Human” consisted of Ben “Merlini” Wu, William “Blitz” Lee, Ioannis “Fogged” Loucas, David “MoonMeander” Tan and Austin “Capitalist” Walsh. Each one of these players is ranked in the 99.5th percentile of players globally. Despite this, Open AI still made them look like incompetent opponents.
Open AI was introduced at The International 7 and was showcased in a one versus one mid only setting. In these games the bot played against some of the best DotA 2 players in the world. Like the show matches that were just played, the bot won convincingly almost every time. Danil “Dendi” Ishutin, who many consider the face of DotA 2, was absolutely manhandled by the bot. Many people doubted the ability of the developers to transfer this success into a team oriented play style. However, the Open AI team has silenced those doubts, and has proven that the software can only improve going forward.
Key doubts included an inability to form strategies that could compete against organized teams. These strategies include understanding power spikes between one bot and the team as a whole. For example, competitive teams will determine the best time to fight based on the completion of an item. People were unsure whether or not a team of bots could have the communication to execute during these important times. Furthermore, the community could not fathom how bots could craft and execute strategies during the draft phase. Drafting is very dynamic and based upon the other teams picks so counter picks could be problematic.
Open AI had a pretty straight forward strategy. They chose to use survivable heroes that would not rely on winning their lanes. These heroes included Witch Doctor, Gyrocopter, and Crystal Maiden. They would abuse player mismatches by grouping up and killing solo heroes then use their early advantage to snowball through towers. They would use immense pressure in order to choke out the human team, restricting their ability to come back. Additionally the bots would use a lot of regen in lane to reduce the amount of time in between fights. The strategies employed by the bots very much resemble those seen in The International 4. Using very strong early game heroes to “Deathball” through the other team.
Open AI Weakness
Weaknesses in Open AI were exposed though in the third show match with the human team. The exhibition was only supposed to be a best of three, and the bots won the first two games of the series convincingly. However a third game was played just for fun, but with a little twist. Instead of the bots being allowed to draft their own team, their heroes were chosen by twitch chat and the crowd. The resulting draft gave the bot team a 3% chance of winning the game as it mostly consisted of core heroes. These heroes went against the strategies employed in their previous matches. They required items and levels to fight competently and without that farm they were sitting ducks. And although the human victory was much more delayed than the bots, they still secured the victory.
This match was the most interesting of the three because it allowed the viewer some insight into how the bots operate. Initially they had four heroes in the bottom lane and were able to get a few kills to start the game. This is emblematic of the same strategies the bots used before, early pressure.
However they began to panic as the game went on, towers were falling and the human team was farming too much. This is when it was exposed that the bots would do anything in order to protect their towers. The bots would cut creep waves before those creeps were able to make it to a tower in order to slow the human team’s push. While this worked at first, their strategy would eventually fail as the humans exploited the fact that the bots were playing alone. Furthermore, the bots would get caught in odd positions and subsequently die to the human team’s rotations.
It is important to note that the Open AI software is still in development and there are some obvious issues with it. For example the hero selection is limited to only 18 of DotA’s 112 total heroes. Furthermore the bots are still incapable of courier management, and as a result they are allowed 5 couriers for each individual bot at the beginning of the game. They are also very limited in their use of support items such as Smoke of Deceit and wards which are often misused. But this show match gave much promise to AI’s ability to beat professional players in every facet of competitive play. As the software continues to learn by playing itself and others, it can only improve which is a slightly scary prospect.
Featured image courtesy of Open AI
“From Our Haus to Yours”