Trust, but verify: Model-Based exploration in Sparse Reward Environments