Before we start off, congratulations to team Liquid for winning the most prestigious DotA2 Tournament, The International 2017 (much love to my boy Miracle). However, we are here to talk about the presence of an incredible Artifical Intelligence (AI) which puts even the best players of DotA2 on their knees in a duel of mano-a-mano battle between 2 Shadow Fiends, a character known for its capabilities and popularity for such duels. It is deemed as not being able to make any mistakes, which is definitely not true since the AI lost once (and only once) to a player named Pajkatt. The AI is developed by OpenAI as a project to test the capabilities of the AI.
OpenAI is a startup company driven in a passionate will to create Artificial General Intelligence (AGI), which is going to be the next “big thing” besides Big Data and Half Life 3. OpenAI was sponsored by influential people such as Y Combinator president Sam Altman and Business magnate/Tesla Founder Elon Musk. Keep in mind that this company is non-profit, so kudos to all of the company’s personnel on doing an extremely good job without thinking about profit (their researches are secure though. Imagine getting money from The Elon Musk).
So how does the AI actually learn how to play DotA2 and beat the professional players of today?
The first test happened on March 1st, which involved the AI winning over Arteezy, a top US player. The AI utilised a method of learning called machine learning. Machine learning is the method which allows computers to run a certain sequence of tasks, detect the mistakes and anything remotely wrong in the run, and fix them. OpenAI’s DotA2 AI has been tested many times starting from playing with hard-coded bot (which basically acts in the same behavior for every situation). It will then learn the basics, which continues further to more advanced mechanics and strategies. When the AI’s prowess was displayed in TI7, it was able to (pardon the jargons) manipulate creep aggro, creepblock, and fake a shadowraze, all of which are advanced strategies which many human players cannot fully use in their games.
Similar machine learning application is done on a computer program made by DeepMind, a Google AI developer. Recently, the particular Google’s AI was given an avatar/body of a human ragdoll and told to go through a sequence of points, one point to another. Unfortunately, that is all the AI was told to do, without any explanation about how to move. As such, it seems impossible for the AI to navigate itself through the terrain (which has gaps and different heights of platform, mind you) when it had not even learned to even walk. The AI was programmed well so that it utilised machine learning, which allowed it to learn to move, run, and even jump after some period of time.
These kinds of AI are not only utilized in running through obstacles and beating professional players in their games. Some common technology which we take for granted almost every day is advertisements you encounter when you browse. Your browser can detect your browse pattern and habit. Advertisements will pop up with relevance similar to web pages that you visit often. Similar application is done on spam detection in your emails.