Inquiry Regarding Discrepancies in Results from Your Published LA3P Code

[Franklin-19]

Hello, thank you for sharing your work and the code for LA3P!

Our team is working on improving the TD3 reinforcement learning algorithm and using your implementation as a baseline for comparison. However, we've encountered some issues replicating the results presented in your paper. Specifically, under both V2 and V4 conditions, our results are significantly lower than expected.

For instance, in the HalfCheetah environment, the agent's maximum cumulative return is only around 3,000, and in the Ant environment, the performance stays around 2,000, sometimes even falling behind TD3. We're puzzled by this and wondering if there were any minor code modifications not reflected in the current repository that could account for the higher performance in the published results.

Any insights or clarifications would be greatly appreciated, as we plan to cite your approach in our upcoming paper. Thank you in advance for your help!