Behaviour-Conditioned Policies for Cooperative Reinforcement Learning Tasks

The cooperation among AI systems, and between systems humans is becoming increasingly important. In various real-world tasks, an agent needs to cooperate with unknown partner types. This requires the assess behaviour of during a cooperative task adjust its own policy support cooperation. Deep reinforcement learning models can be trained deliver required functionality but are known suffer from sample inefficiency slow learning. However, adapting ongoing ability type quickly. We suggest method, where we synthetically produce populations agents different behavioural patterns together ground truth data their behaviour, use this for training meta-learner. additionally architecture, which efficiently generated gain meta-learning capability. When equipped such meta-learner, it capable quickly types in new situations. method used automatically form distribution meta-training emerging behaviours that arise, example, through self-play.