Deep learning is an emerging field that has shown many recent breakthroughs due to advances in computing power, algorithms and the scale of data sets. Within reinforcement learning, deep learning has shown to be effective in game-like environments.
The purpose of this study is to investigate the effectiveness of deep reinforcement learning on Mini World of Bits, a novel benchmark developed by OpenAI. This benchmark serves as a starting point for reinforcement learning approaches to website interaction, which, if mastered, would strongly signal development towards automating tasks on real-world websites.
Ultimately, being able to learn sophisticated sequences of tasks required by browser interaction is a marker of progress towards general artificial intelligence. To our best knowledge, this is the first reported time deep reinforcement learning has been applied to the new datasets.
We look at adapting and evaluating Deep Q-networks and Policy Gradient algorithms for this benchmark. Towards the later stages of training, we are able to see a 10% increase in reward received by our Double Dueling Deep Q-network compared to a baseline random agent