Random Early Detection (RED) is widely deployed as a congestion control mechanism in the routers. However, RED is often limited by the difficulty of tuning its parameters under different congestion levels. Adaptive RED (ARED) addresses this problem by adaptively adjusting the aggressiveness of RED setting to keep the average queue size between minimum threshold (minth) and maximum threshold (maxth). We have observed that the Additive Increase Multiplicative Decrease (AIMD) approach used in ARED can always maintain high link utilization but not effective enough in preventing high packet loss rates under heavy congestion. In this paper, we propose an adaptive version of RED which applies non-linear power function instead of linear AIMD approach to achieve better link utilization as well as minimizing the packet loss rates under various traffic load conditions.