EA is testing self-learning AI agents in Battlefield 1

EA is testing the use of virtual players in Battlefield 1 who teach themselves basic skills almost from scratch, using artificial intelligence. The goal is to eventually use the self-learning agents in games.

The agents are not thrown in at the deep end from scratch: EA first had them analyze the game of human players for half an hour as part of imitation learning. They were then able to train for six consecutive days against old-fashioned bots on multiple machines, bringing the total training time to 300 days. For this, the researchers used temporal difference reinforcement learning.

EA had to simplify the display for the agents and the complexity of the game meant that they were not quick to learn and could ‘think’ little ahead. But they did master the basic skills and were able to react appropriately to situations. For example, if they were low on ammunition, they would look for it. They still lost a game against human opponents and they also sometimes showed strange behavior such as spinning in circles when no target was in sight.

Behavior examples: red: looking for ammunition, blue: patrolling, yellow: looking for places to find something wat

“Battlefield is about much more than defeating the opponent. It involves a lot of strategy, things like teamwork, knowing the map and being familiar with individual classes and gear,” explains Magnus Nordin, who works on AI at EA. as part of the wider Search for Extraordinary Experiences Division, or SEED.

Nordin thinks intelligent AI agents will become part of games in the future where they can adjust their behavior based on interaction with human players. He also thinks that the virtual players will be able to beat human players in a simple game mode within a few years.

Several parties are engaged in research into artificial intelligence in games. For example, DeepMind and Blizzard are testing AI techniques with the rts StarCraft II and OpenAI is training artificial intelligence with Portal, Red Alert 2 and Shovel Knight.

The SEED researchers have published their findings in a paper titled Imitation Learning with Concurrent Actions in 3D Games.