The optimistic exploration value function

The paper presents an approach that uses optimistic initialization and scalarized multi-objective learning to facilitate exploration in the context of model-free reinforcement learning. In contrast to existing optimistic intialization approaches, the approach introduces an extra value function, which is initialized optimistically and then updated using a zero reward function. Linear or Chebyshev scalarization is then used to compound this function with the standard task-related value function, thus forming an exploration policy. The paper concludes with evaluation of the approach on a benchmark task.