📜

Disambiguation of the world’s inventors

Serial inventors often patent under different names, firms, places, and time periods. Identifying the correct individuals is a challenging process and remains to be done for the inventors of the world contained in PATSTAT. The goal of this project is to address the ambiguity issue of inventors in the PATSTAS via a combination tool of ML and BERT. Promising results have been illustrated on a small dataset and the large dataset is currently in progress.

Exploring extensions of Qmix

QMIX is a state-of-art value-based deep Multi-Agent Reinforcement Learning (MARL) algorithm in the paradigm of centralized training with decentralized execution (CTDE). However, experimental results show that QMIX performs unstably on some cooperative multi-agent tasks. Therefore, to improve the performance and stability of QMIX, we explore several extensions of QMIX and proposed d-QMIX, which takes dueling network architecture to compute the individual value function. Experiments show that d-QMIX significantly outperforms existing well-known value-based MARL methods in the paradigm of CTDE on different cooperative tasks.