Hamburg University of Technology (tuhh) | intelligent problem solving as integrated hierarchical reinforcement learning

【 title 】Intelligent problem-solving as integrated hierarchical reinforcement learning

【 The author team 】Manfred Eppe, Christian Gumbsch, Matthias Kerzel, Phuong D. H. Nguyen

【 Date of publication 】2022.1.25

【 Thesis link 】https://www.nature.com/articles/s42256-021-00433-9?utm_source=xmol&utm_medium=affiliate&utm_content=meta&utm_campaign=DDCN_1_GL01_metadata

【 Recommended reasons 】 According to cognitive psychology and related disciplines , The development of complex problem-solving behavior of biological subjects depends on hierarchical cognitive mechanism . Hierarchical reinforcement learning is a promising computing method , Eventually, similar problem-solving behaviors may occur in artificial intelligence and robots . However , up to now , Many human and non-human animals have significantly better problem-solving abilities than artificial systems . This paper proposes the steps to integrate the biologically inspired stratification mechanism , To realize the advanced problem-solving skills of artificial intelligence agent . This paper first reviews the literature of cognitive psychology , To emphasize the importance of combinatorial abstraction and prediction processing . then , Link the insights gained to contemporary hierarchical reinforcement learning methods . The results show that , All identified cognitive mechanisms have been implemented separately in an isolated computing architecture , This raises a problem , That is why there is no single unified architecture that integrates them . As a final contribution , This article solves this problem by providing a comprehensive view of the computational challenges of developing this unified architecture . It is hoped that the results of this paper can guide the development of more complex cognitive inspired hierarchical machine learning architecture .