Mentor: Ishika Singh
Project Members: Aditya Gupta, Arka Das, Padmaja Chavan, Sajal Goyal
Atari 2600 is a challenging RL testbed that presents agents with a high dimensional visual input (210 x 160 RGB video) and a diverse and interesting set of tasks that were designed to be difficult for humans players. The network is not provided with any game specific information or hand-designed visual features, it learns just from the video input, the reward, terminal signals, and the set of possible actions. The goal is to connect a RL algorithm to an deep neural network which operates directly on RGB images and by using stochastic gradient updates. Utilizing experience replay which is applying Q-learning updates to samples of experience, random from the pool of stored samples. Although RL takes a lot of time to train, it is capable of doing tasks which are impossible by pure deep learning.