Reinforcement Learning in Creative Skill Development

Reinforcement Learning (RL) is a very effective computational model that has been used to model adaptive decision-making but not exploited yet in advancing development in creative skills. This paper explores the application of RL in developing creativity in areas of visual art, music composition, design, and writing. We can place RL as a natural process of directing agents to new and valuable outcomes by theorizing creativity as a process that can be learned, and improved through exploration, evaluation and refinement. The study incorporates the knowledge in the areas of cognitive science, computational creativity, and available applications of RL to develop a methodology (environment simulation, creative dataset, and reward-based learning algorithm) to achieve environment simulation. We have an RL-based creative agent, which can interact with environment-specific domains via the feedback loop and dynamically determined reward functions that encourage originality, coherence and aesthetic or usability. The model structure focuses on multi-modal input representation, hierarchical acquisition of policy and adapting the reward modulation in order to stimulate cognitive diversity and intentional exploration. Testing is based on quantitative measures of creativity (novelty support, distributional distance, and statistical surprise) and qualitative measures of creativity, as determined by human subjects. Findings show that creative heuristics can be gradually learned by the agents of RL and that more and more original artifacts can be produced by the agents and that the agents update their strategies based on the evaluative feedback

Cognition · Computational Creativity · Computational model · Creativity · Heuristics · Reinforcement learning · Artificial Intelligence in Games · Creativity in Education and Neuroscience · Music Technology and Sound Studies

• Machine learning and deep learning

  Open Access•Christian Janiesch, Patrick Zschech et al.•Electronic Markets•2021

  Add to list