Challenges in browser automation
Potential practical applications of browser automation with RL are attractive but have one very serious drawback: they’re too large to be used for research and the comparison of methods. In fact, the implementation of a full-sized web scraping system could take months of effort from a team, and most of the issues would not be directly related to RL, like data gathering, browser engine communication, input and output representation, and lots of other questions that real production system development consists of.
By solving all these issues, we can easily miss the forest by looking at the trees. That’s why researchers love benchmark datasets, like MNIST, ImageNet, and the Atari suite. However, not every problem makes a good benchmark. On the one hand, it should be simple enough to allow quick experimentation and comparison between methods. On the other hand, the benchmark has to be challenging and leave room for improvement. For...
