This was an interesting competition organized by Iran's national elites foundation to build a Persian language AI assistant. We won this competition as the DeepAid team (with Ali Amigh and Amir Amirnezhad), where I was working on the software part of the project, and Ali and Amir did an excellent job on the hardware part. In order to improve our model I designed a Telegram bot which helped us to debug and collect more data. Below you can see the implementation of the Classifier and NER models which are the core of this assistant and are based on ParsBert transformer model.
This is how I implemented a multi-agent Q-Learning problem using Tensorflow-agents library which for overcoming the rewarding issues I passed the rewards as a part of the observation (state). The environment is defined as a 20x20 grid which 4 agents (animals) can take one of the 8 possible actions: