Skip to content

Latest commit

 

History

History
32 lines (19 loc) · 2.61 KB

week-04-and-05.md

File metadata and controls

32 lines (19 loc) · 2.61 KB
Raw

Week 4 and 5

8th June - 21st June , 2020

Work done

  • Added a common function(agentTest) for tests in QLearning.
  • Added C51's layout as a separate network in q_networks, with Predict and Forward functions completed.
  • Added the calculation for target distribution in C51, along with the Backward function.
  • Made separate training functions for categorical and non-categotical versions of dqn, thus ensuring maximum reuse of code.
  • Initially I thought of creating a separate replay method for N-step, but later found a way to addon changes to RandomReplay and PrioritizedReplay. Now multi-step can be enabled by just passing n as a last parameter, while initializing the replay method.

Problems faced

A rather interesting issue that got unnoticed was the way noisy layers were addressed in noisy implementations of simple and dueling dqn. The layers had a special function (called ResetNoise) to be called after each update step. So, we had to somehow refer to all the noisy layers for that. For this, we were using a list of Layer pointers. This was a problem, since when we copied over the learningNetwork to targetNetwork in q_learning, both the networks referred to the same noisy layers. This bug was resolved by storing just the indices where the noisy layers were used, and calling ResetNoise on each noisy layer using boost::get<>.

Apart from that, the target distribution calculation for C51 dealt with a good amount of matrix manupilations, which got a bit confusing in the begining, but was very satisfying in the end, when it got over.

By the way, the c51 implementation doesn't converge. I searched a lot, but couldnt find what the error was. I even made a pytorch implementation to compare the algo with. The same workflow converges on pytorch, but doesn't in mlpack. We are still investigating what the cause might be. Here's the link.

Lastly:

By now, I'm pretty sure that most colleges, especially ours, are going to conduct their next semester completely online. I guess it'll be a new experience altogether! I'm not sure when's the next time we are going to see our college...

This week, I saw two great movies, both incredible in their respective genres.

The 'Green Mile' had a very profound effect on me.

'Knives Out' is a recent film, and I was surprised why I hadn't heard of it earlier, considering the above average ratings and familiar artists. Anyways, I guess it can be counted among the best mystery movies that I've seen.

Btw, I recently found out that watching unboxing videos can feel quite satisfying sometimes..

Thanks for reading, and have a good one ;)