An Adaptive Deep RL Method for Non-Stationary Environments with Piecewise Stable Context