2
$\begingroup$

I've decided to make my bachelor thesis in RL.

I am currently struggling in finding a proper problematic. I am interested in multi-agent RL with the dilemma between selfishness and cooperation.

I only have 2 months to complete this and I'm afraid that multi-agent RL is too difficult and I don't have the knowledge and time to nicely learn this topic.

Do you have any problematics for a bachelor level student ? I've done a tiny Q-learning algorithm to solve Open-AI text based environements.

Thanks in advance for the help !

| improve this question | |
$\endgroup$
  • $\begingroup$ Hi Romain! This website is not suited for recommending career paths. See: ai.stackexchange.com/help/on-topic. I would encourage you to reformulate your question, so that it is on-topic. $\endgroup$ – nbro♦ Apr 9 '19 at 22:15
  • $\begingroup$ @nbro This is not a career path recommendation. This is asking "what are some simple open problems in multiagent RL". This question is definitely on topic, although it might be difficult to answer. $\endgroup$ – John Doucette♦ Apr 13 '19 at 0:55
3
$\begingroup$

Welcome to AI.SE @Romain B.

I have several undergraduates working on multiagent deep RL problems for their theses, but most of them have been working for 8-9 months. 2 might be a stretch.

Good multiagent deep RL problems for a bachelor's thesis might look something like:

  1. Pick an older video game, which has been studied using Deep RL, but not in depth. Right now my students have been liking Nintendo 64 games.
  2. Read the papers that study this game already.
  3. Pick one of the described approaches and reproduce the paper's results in your own system.
  4. Pick one of the parameters that the paper does not explore changing, and see what happens as you change it.

This probably does not lead to a publishable result, but it is real science and can make for a fine undergraduate thesis.

A slightly harder project, which may require more time, would be to examine the "future work" sections of these papers, and perform one of the experiments suggested there. These experiments often lead to small publishable results.

| improve this answer | |
$\endgroup$
  • $\begingroup$ Giving direct advice to students in the form “Read a paper” and “pick an older video game” are not helpful because the domain of writing an academic paper is way to complex for such easy hints. The better idea is to provide scenario based explanations. For example “if a paper is about a video game, then the danger is high, that it won't be perceived as science”, or “if the results of a previous work is reproduced in the own paper, then the task of writing the paper is much easier”. $\endgroup$ – Manuel Rodriguez Apr 13 '19 at 7:12
  • $\begingroup$ @ManuelRodriguez I have to disagree about its helpfulness. It really is not that complex for a typical senior student who has already taken an AI course or three to do what I've just described. The whole point of a typical honors curriculum to produce students who, at the end of a 4 year degree, really can just go read some papers, extend the results therein, and write their own paper about it. I know this because my own students do this every year, on the basis of similar suggestions. $\endgroup$ – John Doucette♦ Apr 13 '19 at 14:24
  • $\begingroup$ Thank you for your answer, I chose generalization in grid-world environments :) $\endgroup$ – Romain B. Apr 30 '19 at 8:24

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.