Constrainedepisodicreinforcementlearningin concave-convexandknapsacksettings