Multi-step DDQN - Githubissues

lgooo / SUMO-RL-Coverage

We provide an open source software package for AV based simulation and testing running a docker container

0 stars 0 forks source link

Multi-step DDQN #24

Closed kihyukh closed 2 years ago

kihyukh commented 2 years ago

Currently, only a single step (s, a, r, s') is considered for training DDQN. There is empirical study that multi-step training performs better: https://rayyoh.github.io/files/2017-Rainbow.pdf

Let's implement multi-step DDQN version.

Part of the code affected:

Experience replay structure
Loss function

lgooo commented 2 years ago

So far, multi_steps = 3 produced the best results.