Multi-agentactiveperceptionwithpredictionrewards