Artificial Intelligence (AI) is affecting the way we work, learn, shop, and now creating new opportunities for playing and watching our favourite sports.
Using AI and machine learning to mimic the behaviour of the likes of Cristiano Ronaldo and Lionel Messi, a team of researchers from the at AV俱乐部 were recently named as runners up in 2020鈥檚 largest international AI soccer simulation competition, . This is the first time a Canadian team has made the finals for more than 10 years. 聽
International robotics competition RoboCup uses soccer simulation to promote robotics and AI research with the research findings used to advance many areas. By 2050, the competition aims to train a team of fully autonomous humanoid robots to win a soccer game against the winner of the most recent World Cup.
A new way of using AI
RoboCup uses a variety of AI and machine learning models to train autonomous 鈥榩layers鈥� in different positions to collaborate and play together in an intelligent way.
鈥淪o, most of our decision making is based on AI,鈥� says Mahtab Sarvmaili, a PhD student in the Faculty of Computer Science. 鈥淚f you don鈥檛 use it, the number of decisions is too much. We have to use AI to have a better approach, or better way of playing the game. Now we are trying to make it more similar to the real world, especially by using the reinforcement learning method that鈥檚 like training the agent without having any actual supervision from humans.鈥�
Stan Matwin, Canada Research Chair and director of the Institute for Big Data Analytics adds, 鈥淭his is a new way of using AI where we use a whole range of AI methods to create a simulation of perception so that the agents or 鈥榩layers鈥� work autonomously but also learn how to collaborate.鈥�
Addressing bigger issues
While these methods are presenting new ways to play and consume sports particularly during a time where much in person activity is suspended, the team is identifying other ways to use this research to address bigger issues such as cleaning our oceans.
鈥淢ainly what can be extracted from this is how agents or robots can collaborate to achieve a certain goal,鈥� explains Amilcar Soares, an assistant professor at Memorial University and former AV俱乐部 postdoc. 鈥淟et鈥檚 think about the ocean domain. One interesting problem that I see several companies discussing is how we can make autonomous agents collaborate to clean the ocean. So, you could have multiple agents and humans can assign tasks to the agents to collaborate in a way that the final task will clean the ocean or river.鈥�
Mahtab echoes that this method of training agents makes them learn behaviours, resulting in them working in the most effective and intelligent way, be that the fastest or by using the least amount of energy. This can also be translated to the future of video games.
鈥淚f you consider these very complex games, such as Warcraft, they are not fully autonomous, they are controlled by a central server and the server tells the agent what to do,鈥� she says. 鈥淲hat we are doing here is way more complex because we are distributing the decision making and each one of these players acts like a human and thinks about the game. So now, if they learn how to collaborate with one another in this complex environment that is constantly changing, we can apply this model on the other video games and make them more realistic and interesting.鈥�
Value of human experience
AI and Machine Learning may have controlled much of the outcome for the team during the RoboCup competition, but they argue that human experience and knowledge is still essential, and they took some inspiration from the best.
鈥淣ader (Zare, chief architect of the RoboCup team) has developed a defensive strategy trying to mimic the strategy of Barcelona FC. So, we are trying to get an idea from the real-world players, and optimize their behaviour,鈥� says Mahtab. 鈥淚t鈥檚 not actually like them, but we want to make our team close to the real human players.鈥�
鈥淚t鈥檚 important to point out that during the actual competition we evaluated that it was taking very long for us to make a shot at the goal.鈥� Dr. Soares expands. 鈥淲e checked the threshold for shooting and we saw that it was very low. This was an observation by a human that at some point, even though we had so much data to train, just using the machine it was not perfect, it needed these human interventions to play better.鈥�
On the flip side, he argues that humans could learn more about the game from the technology and that this could change the way sports are played and watched forever.
鈥淲e could have, for example, a virtual assistant coach that could help the human coach and say the strategy isn鈥檛 working, why not replace these two players etc.鈥� Dr. Soares says. 鈥淭his could be learned from a platform like that in a way that this decision making can be provided to the coach. I think this could be the future. Even right now it鈥檚 possible, it鈥檚 a matter of just modelling properly.鈥�
Nader adds, 鈥淚n real football, we can use knowledge from RoboCup to analyze a real soccer game. We can use this to predict the result of games and to show the benefit of doing certain things with different positions.鈥�