I have had handfuls of university students come to me for help in the full pocess of deciding what to make a paper about, coding the architectures for the paper, and reviewing the paper. That includes masters papers,and it has sorely affected my opinion of formal educations ability to teach DRL and its reputation.
Most Ml, and especially DRL papers are useless, or worse, lies. I would say over 60% of them. The authors are usually motivated by graduation, or social situation, and often contain faked results. This is especially true of deep RL papers. Deep Reinforcement Learning is hard.
You cant just fake the data of the model succeeding because the paper is due (Saw this three times, fought them). Why are you using a discrete action space for a continuous environment?? (Usually its because they cant get the code to work for non Q based algorithms. Until PPO came out the entire Actor Critic family was really hard to converge, and everyone was failing with all the variants of DDPG. After PPO that changed, but the average ml paper writer couldnt succesfully implement it. I failed a few times, took me two solid weeks of hacking at it. Students who fail to implement their AC family algorithm would go use DQ from the 2012 Mnih. paper because they learned Q learning on gridworlds in class, but it only has discrete outputs... not only that but DQ was replaced by DDQ in 2015, and under no circumstances other than education should the original DQ be used. Real DRL just has almost nothing in common with the gridworld and Tabular Q based classes people are taking in school.) Why are you truncating the model outputs instead of letting the channel between two models be in feature space, despite not using those truncated outputs as some sort of gradient constraint? (Many companies are still doing this) Where is the epsilon decay schedule in your paper?? and basic entropy graphs, convergence graphs... you cant just publish the env score. you should ensure convergence on a control environment first. No you cant continuously shape the rewards manually throughout the paper to raise performance. No your model wont converge when your environment has sparse rewards or prior gates. No you cant start your discount factor at zero. (lol)
Some of the university students I assisted were absolutely horrible at programming, gave up early and lied in the results. Now I dont have a problem with people practicing and learning but it was clear to me that this wasn't the right setting and they didnt have the right resources or oversight to succeed, and that sucks.
But people see their paper anyways, assume its valid, and they got a degree that year, so I guess I better bow to them?
Whats your criteria for expert? Being in the top 10% of DRL practitioners is a fair measure. I was certainly in that fraction at the peak of my study. Relative to 99.99% of the world population, and 90% of the people who even know what DRL is, I am an expert. I am an expert in Deep Reinforcement Learning. If you have any paid projects you need help with let me know.