Maze Solver: A Federated Reinforcement Learning Approach
Aayush Sharma, Harjeet Kaur, Deepak Prashar · 2023
In today’s era where Machine Learning and Deep Learning are household names and Artificial intelligence is rapidly evolving people are neglecting a basic issue in this type of setting the privacy issue. The Machine Learning and Deep Learning models need a large amount of data to train and while they use the data from the internet, they also use the data collected from the users. All the data used for the training of such models are saved on a single server prone to lots of cyber-attacks such as SQL injection, DDOS, Data leakage, etc.… In order to solve this issue in 2016 Google released a new way to train a Machine Learning or Deep Learning model this approach is known as the Federated Learning letter Google implemented the same framework on a Google Keyboard to create a next-word prediction model without saving any user data on their central server. This framework proved to be a privacy-preserving as well as performance-boosting framework. In order to harness the power of multiple IOTs and edge devices in rather computational resources-hungry tasks such as reinforcement learning researchers started to move the focus to Federated Reinforcement Learning. In this research work an implementation of Q learning a reinforcement learning technique is done in a Federated Environment. Letter the results from both Q learning and Federated Q Learning are compared and analyzed to find out the advantages Federated Reinforcement Learning has over centralized Q Learning or any reinforcement Learning technique.