Palo Alto, Calif., July 10, 2018 — Maana, the pioneer of digital knowledge technology, today announced its researcher Dr. Fangkai Yang will present research paper “PEORL: Integrating Symbolic Planning and Hierarchical Reinforcement Learning for Robust Decision-Making”, co-authored with Maana chief scientist Dr. Steven Gustafson, Prof. Bo Liu and Daoming Lyu from Auburn University, in a technical sessions at the 27th International Joint Conference on Artificial Intelligence (IJCAI) and the 23rd European Conference on Artificial Intelligence, the premier international gathering of researchers in Artificial Intelligence, to be held from July 13–19 in Stockholm, Sweden.
What: Research paper “PEORL: Integrating Symbolic Planning and Hierarchical Reinforcement Learning for Robust Decision-Making”
Who: Dr. Fangkai Yang, Researcher at Maana
When: July 17, 2018, 14:55
Where: 27th International Joint Conference on Artificial Intelligence (IJCAI) and the 23rd European Conference on Artificial Intelligence in Stockholm, Sweden, Room C7
IJCAI is the most established premiere academic conference on artificial intelligence, starting from 1969. IJCAI-18 is part of the Federated AI Meeting that takes place at Stockholmsmässan in Stockholm July 9-19. Other conferences include AAMAS, ICML, ICCBRand SoCS. The World Computer Chess Championships will also take place in parallel. More than 5,000 researchers, technologists and experts are expected to attend IJCAI and its partner AI conferences. Keynote speakers include Prof. Yann LeCun (Facebook), Prof. Max Tegmark (MIT), Prof. Joshua Tenenbaum (MIT), and many more. The main conference will hold 800 separate AI seminars, industry days, demonstration sessions, robotics showcases and exhibition.
During the conference, Dr. Yang will present the latest advancement of its theoretical study underpinning the Maana knowledge platform. The research paper describes a unified framework that integrates knowledge-based artificial intelligence, in particular, knowledge representation and automated planning, with reinforcement learning. This framework allows the agent to generate its optimal behavior by reinforcement learning from the interaction with the environment, guided through reasoning and planning with explicitly represented domain knowledge. This process enables planning and learning to mutually benefit each other so that the behavior rapidly converges to the optimal for complex and dynamic domains.
This research supports the user-guided, machine-assisted decision-making methodology of the Maana platform, where human knowledge helps the intelligent agent to automatically generate its solutions for business problems and proposes recommendations to the user. The agent can further improve its recommendation by observing the effectiveness of its solution or directly receiving user feedback. Eventually, the agent can reach optimal decision-making, utilizing both explicitly formulated human knowledge and personalized feedback, for different users and different problems.
The theoretical fruit of this paper has already led to the development of the AutoML service. AutoML service automatically recommends top performing machine learning pipeline and hyper-parameters given a particular dataset. Different from other AutoML system that mainly focuses on automating ML tasks using a black box, AutoML service allows for extracting knowledge learned from the dataset, a step towards “interpretable AI”. The paper, entitled “Program Search for Machine Learning Pipelines Leveraging Symbolic Planning and Reinforcement Learning” was presented by Dr. Gustafson in 16th Genetic Programming Theory & Practice (GTPT XVI), in May 2018.
For more information about IJCAI, visit: https://www.ijcai-18.org/
Maana organizes the world’s industrial data and human expertise into digital knowledge to enable millions of experts to make better decisions affecting the global economy and human civilization. Using our Knowledge Platform™ and its proprietary Computational Knowledge Graph™, Fortune Global 500 industrial companies drive answers to operational questions that direct the future. Operating internationally, Maana is a privately-held company with offices in Menlo Park, California; Bellevue, Washington; Houston, Texas; London, UK; and Dhahran, Saudi Arabia.
Stay in the know with the latest information about Maana services, events, news and best practices by email.