So, I’m looking to automate some tasks at my job. I work at an engineering company. One of my tasks is produce these “reports” in excel that track some design metrics in our company. It’s a soul sucking task. It’s not particularly hard to do, but there are thousands of these reports to do. I’d like to employ an assistant to help me. With that said, I read this article below and it gave me an idea of doing something similar.
I read this article here on something OpenAI did: https://venturebeat.com/2019/09/17/openai-and-deepmind-teach-ai-to-work-as-a-team-by-playing-hide-and-seek/
They used the term “Undirected exploration”, where the agents evolved there understanding of the world around them by themselves and eventually completed the task at hand using their own methods.
I was thinking along the same lines, where I could employ the agent within the environment of my specific Excel template, give it full control of the Excel API (VBA), access to the needed data and see if it can work towards creating an accurate report in the required format. I don’t care how it does it, as long as the end result is close to what I need. The data it would need to produce ranges from images, to string data and of course numbers. All in a specific format and most importantly, mathematically correct.
I’m not really sure what kind of network will be suitable here, or if I’ll need multiple networks. My initial thoughts would be to use an RNN, and because the agent would be unsupervised, reinforcement learning might help it to reach its goal. I say RNN because I’m envisioning the task as a sequential one. Several steps are needed to accomplish the task, and memory of the last step taken might be needed.
In any case, this will be largely experimental. I don’t know if it will work, or what kind of conclusions it will come too. I’m hoping for the best.
My question is, what kind of network(s) do you recommend for this type of undirected exploration and learning?