Exploring the relationship of reward and punishment in reinforcement learning

We present a reinforcement learning algorithm based on Dyna-Sarsa that utilizes separate representations of reward and punishment when guiding state-action value learning and action selection. The adoption of policy meta-learning optimized by a genetic algorithm is explored and results in the context of a two-armed bandit goal-navigation task in a simple grid world are presented. The findings argue for an important role for a genetic algorithm approach for constructing the foundations of autonomous reinforcement learning agents.