Pittsburgh University researcher uses video games to teach an AI algorithm to learn from its mistakes

Study found that the most successful players in the video game used 'Monte Carlo tree search strategy' to generate data and make decisions

A researcher at Pittsburgh University is using computer games to try and teach an artificial intelligence algorithm to learn from its mistakes.

According to Dr Daniel Jiang, assistant professor of industrial engineering at the University of Pittsburgh's Swanson School of Engineering, the aim of the study is to devise an algorithm capable of learning from its mistakes.

In the study, Dr Jiang used multiplayer online battle arena (MOBA) video games to test the effectiveness of his algorithm. In a MOBA game, such as DotA 2, Heroes of the Storm or League of Legends, players have a mission to raze the opponent's base while protecting their own base.

The game offers a ‘hero' character that players can use/control to accomplish their mission.

Dr Jiang created an algorithm that could evaluate 41 pieces of information in a MOBA game and produce an output that was one of 22 different actions (including attacks, movement, and special moves).

The study compared different training methods used by various players and found that the most successful AI player used what is called the ‘Monte Carlo tree search' strategy - a heuristic search algorithm - to generate data and make a decision.

The Monte Carlo tree search strategy is widely used in Go, Chess and Shogi, and has even been implemented in the popular Total War: Rome II strategy video game, according to its maker Creative Assembly.

According to Dr Jiang, his algorithm was able to analyze the game results to give more weight to actions that proved successful.

After multiple iterations of the game, the algorithm learnt to persist with the more successful actions, thereby increasing its chances of winning the game. The results also demonstrated the effectiveness of the Monte Carlo tree search in training an agent to succeed at making complex decisions in real-time.

The results of the study were presented at the 2018 International Conference on Machine Learning in Stockholm, Sweden this past summer.

The Computing Technology Product Awards 2018 are coming!

Recognising the very best in technology talent, the Awards will recognise innovation and technology excellence across the range of mobility, software, infrastructure, security, cloud computing - and people.

To find out more, to enter and to reserve your table at the Awards on 30 November, check out our dedicated Computing Technology Product Awards website