Reinforcement learning stuck on the cluster

I'm working on a reinforcement learning project with a complex environment that requires a significant amount of time to complete, which has led to a slow learning process. When running the code on my local machine using the parallel computing toolbox with three cores, it works perfectly fine. However, when I attempted to utilize my university's cluster to expedite the training process, I encountered issues. With a low number of workers (specifically, three cores), the training process gets stuck, as depicted in the first image below. On the other hand, when using a higher number of workers (e.g., 10 cores or more), the training process fails to initiate altogether, as shown in the second image below. I'm uncertain whether this problem stems from the cluster's configuration, a Matlab-related issue, or a limitation within the reinforcement learning toolbox. Any insights on potential resolutions would be greatly appreciated.

[Emmanouil Tzorakoleftherakis]

Can you please create a technical support case? This may be hard to debug otherwise

[Ahmad Momani]

I just submitted one

[Mike Croucher]

Did you get this resolved?

[duoke]

Hello, is there a job submission for this anywhere, or do you just need to run the code normally with parallel pooling turned on?

Answer 1

[Satwik]

Hi,

One reason for the issue you are experiencing could be incorrect configuration of the cluster profile. You can verify that the cluster profile is set up correctly by following the steps outlined in this MATLAB answer provided by the MathWorks Support Team: https://www.mathworks.com/matlabcentral/answers/1590004-how-do-i-validate-my-parallel-cluster-profile-in-matlab.

1. Open MATLAB and go to the ‘Home’ tab in the ribbon.
2. In the ‘Environment’ group, click on the ‘Parallel’ dropdown.
3. From the dropdown, select ‘Create and Manage Clusters...’.
4. In the Cluster Profile Manager window that opens, you may see multiple Cluster Profiles on the left-hand side. Select the one you wish to validate.
5. In the ribbon of the Cluster Profile Manager, click the ‘Validate’ button to initiate the validation process.

This will perform a five-stage validation process, displaying the results in real-time, and should help determine if the issue is related to the cluster configuration.