The field of Artificial Intelligence (AI) has grown by leaps and bounds since the English mathematician Alan Turing first asked the question: “Can machines think?”. His paper, “Computing Machinery and Intelligence”, published in 1950, would go on to pave the way for the field that would eventually come to be known as AI. Since then, AI has made significant impacts in fields such as healthcare, logistics, and finance, among others. Over the last decade, the gaming industry has increasingly become a major testing ground for new AI algorithms. This is because games can provide numeric scores and win-lose outcomes, making performance results easily comparable to humans or other AI. The latest breakthrough in AI-powered gameplay came in April 2019, when OpenAI’s Defense of the Ancients 2 (Dota 2) artificial intelligence program defeated the Dota 2 powerhouse professional team, Team OG, in back-to-back victories at the OpenAI Five Finals in San Francisco, CA.
OpenAI, a San Francisco-based AI research company with notable founders Sam Altman and Elon Musk, has a mission to ensure that artificial general intelligence benefits all of humanity. OpenAI uses Dota 2, a complex strategy and action game, as a testbed for AI systems that captures real-world complexity. The objective of OpenAI’s Dota training system is to show that AI algorithms can learn what to do in complex situations while also considering variables such as teamwork, long-term horizons, and hidden information.
The Open AI Five Finals was a best-of-three competition featuring five top Dota 2 professional players from Team OG and five OpenAI neural network bots. In 2018, Team OG won The International, an annual tournament organized by Valve, which has become the premier Dota 2 event. Due to the complexity of Dota 2 and the near limitless combinations of unique characters (there are over 100), skill trees, and items for players to buy, certain handicaps were implemented to account for conditions and variables that OpenAI hadn’t trained the bots for. These handicaps included:
- Reducing the number of heroes available for selection to 17 for each match;
- The implementation of a “Captain’s Draft” game mode, which allows each team to strategically ban heroes from selection by the other team; and
- The disabling of summoning and illusion capabilities which can introduce additional variables such as unique creatures and player copies.
OpenAI’s bots were trained using reinforcement learning, a trial-and-error approach to generating a deep neural network designed to achieve goals using rewards and punishments. The neural networks were trained using OpenAI’s general-purpose training system, Rapid, which had OpenAI Five play copies of itself a staggering amount of times in preparation for the tournament. Over 180 years of gameplay data was collected, which consumed 128,000 CPU cores and 256 GPUs to generate. While the handicaps illustrate that more work needs to be done in OpenAI Five’s development, it also displays the incredible advances made in a relatively short period of time – OpenAI Five lost a similar best-of-three match against paiN Gaming (eliminated early in The International 2018) in August 2018.
OpenAI’s project shows that while video games have demonstrated their ability to serve as an optimal testing domain for new systems, AI also has the potential to have a profound effect on the gaming industry. One such area is the use of AI software to enhance the training of human players. With the competitive video gaming scene becoming increasingly saturated as new players join competitions and leagues daily, players seeking to increase their grasp of the basics, learn advanced mechanics, and even learn new strategies may turn to AI teachers in the future. SenpAI, a project under development by the company Falcon AI, trains users to become better at certain games through deep learning analyses of gameplay. SenpAI’s Dota 2 beta, which was released in September 2018, has been utilized by close to 20,000 users worldwide. Players are advised by an AI coach on how to effectively attack and defend, and are provided guidance on the potential payoffs associated with alternate approaches in terms of increased (or decreased) winning probabilities. As a result of the positive reception to the beta, SenpAI has initiated development of a League of Legends (LoL) application. As a point of comparison, at the start of 2018 Dota 2 garnered 10.6 million active users while LoL had close to 100 million monthly active users – this increased player base provides an even deeper data pool from which SenpAI’s neural networks can analyze. In addition to AI coaches, AI players are increasingly demonstrating that they can hold their own against their human counterparts. A week after victory over Team OG, OpenAI Five was released publicly via Dota 2’s Arena mode to see how OpenAI would hold up against a diverse player-base of varying skill levels. Between April 18, 2019 and April 21, 2019, OpenAI Five played 42,729 matches (with close to 700 of the matches being played against as many as 1,583 players at once) and had a 99.4% win rate. Despite the Dota community teaming up to find strategies to defeat OpenAI Five, the bots turned out to be much more resilient and difficult to exploit than expected.
How can we help?
Teknos Associates is uniquely positioned to identify and analyze the current and anticipated trends within the Esports and gaming industry. From our significant knowledge of how the Esports and gaming industries have developed, to our first-hand experience providing advisory services to organizations at all stages of development and growth, our team is ready to help brands capitalize on the myriad of opportunities available in the Esports and gaming industries. Our vast experience includes providing valuations, offering structural guidance, and delivering strategic analyses. To learn more, contact Teknos Associates at firstname.lastname@example.org .