Deepmind introduces a new research framework for AI agents in simulated environments such as video games that can interact more flexibly and naturally with humans.
AI systems have achieved great success in video games such as Dota or Starcraft, defeating human professional players. This is made possible by precise reward functions that are tuned to optimize game outcomes: Agents were trained using unique wins and losses calculated by computer code. Where such reward functions are possible, AI agents can sometimes achieve superhuman performance.
But often – especially for everyday human behaviors with open-ended outcomes – there is no such precise reward function.
