search

TesfayZ / CCM_MADRL_MEC

The source code for the paper titled Combinatorial Client-Master Multiagent Deep Reinforcement Learning for Task Offloading in Mobile Edge Computing
57 stars 6 forks source link

Overfitting in reproduction #2

Closed arold66 closed 5 months ago

arold66 commented 5 months ago

I think your work is interesting and have tried to reproduce your work, but I have experienced overfitting many times. I noticed that you mention in your paper that this is normal, but I can't quite understand how overfitting can prove that your algorithm can outperform a benchmark algorithm if it occurs.This is because under the same environmental parameters, sometimes overfitting occurs and sometimes it is able to train and optimise normally.

TesfayZ commented 5 months ago

This depends on the initial weights of the neural networks. The environment is seeded but not the neural network wrights. That's why we run 10 or more experiments and plotted with 95% confidence interval.

On Fri, 19 Apr 2024, 07:53 arold66, @.***> wrote:

I think your work is interesting and have tried to reproduce your work, but I have experienced overfitting many times. I noticed that you mention in your paper that this is normal, but I can't quite understand how overfitting can prove that your algorithm can outperform a benchmark algorithm if it occurs.This is because under the same environmental parameters, sometimes overfitting occurs and sometimes it is able to train and optimise normally.

— Reply to this email directly, view it on GitHub https://github.com/TesfayZ/CCM_MADRL_MEC/issues/2, or unsubscribe https://github.com/notifications/unsubscribe-auth/ANZZ273LNMSYNSPSPM2UWVLY6C5QJAVCNFSM6AAAAABGOTM3K2VHI2DSMVQWIX3LMV43ASLTON2WKOZSGI2TEMRRHAYTSNY . You are receiving this because you are subscribed to this thread.Message ID: @.***>

arold66 commented 5 months ago

This depends on the initial weights of the neural networks. The environment is seeded but not the neural network wrights. That's why we run 10 or more experiments and plotted with 95% confidence interval. … On Fri, 19 Apr 2024, 07:53 arold66, @.> wrote: I think your work is interesting and have tried to reproduce your work, but I have experienced overfitting many times. I noticed that you mention in your paper that this is normal, but I can't quite understand how overfitting can prove that your algorithm can outperform a benchmark algorithm if it occurs.This is because under the same environmental parameters, sometimes overfitting occurs and sometimes it is able to train and optimise normally. — Reply to this email directly, view it on GitHub <#2>, or unsubscribe https://github.com/notifications/unsubscribe-auth/ANZZ273LNMSYNSPSPM2UWVLY6C5QJAVCNFSM6AAAAABGOTM3K2VHI2DSMVQWIX3LMV43ASLTON2WKOZSGI2TEMRRHAYTSNY . You are receiving this because you are subscribed to this thread.Message ID: @.>

Thanks for the answer, I will try to keep reproducing your work, I only tried an experiment with 10 steps per episode before.