About QMix(MG) about on-policy HOT 5 CLOSED

Thanks for the questions! We do evaluate QMIX with a concatenation of the default env global state and all agents' local observations. We do this to study if this helps the performance of QMIX, and to create a fair comparison with MAPPO (AS) which concatenates the global env state with the local agent observation and does not prune features (and hence is a fair comparison with QMIX(MG)). This is seen in Figure 14 in the appendix of our paper. We notice that QMIX (MG) still underperforms MAPPO(AS) on several maps, and is comparable to MAPPO(AS) on the others.

It is important to note that MAPPO's performance with the FP and AS states are very similar; FP significantly outperforms on only 1 map (3s5zvs3s6z) - this is seen in Table 1. Our primary purpose with conducting FP experiments was that reducing the critic's input dimensionality by dropping redundant features can further improve MAPPO's performance - however, this improvement is marginal. You're right that we could have evaluated QMIX by pruning redundant features, but we felt that because we already compared QMIX(MG) with MAPPO(AS), these experiments wouldn't yield many new conclusions.

from on-policy.

And could you please provide the code of QMix(MG)? Thanks!

from on-policy.

In my opinion, the pruning operation itself is a form of prior human knowledge. However, only MAPPO can take advantage of this knowledge under the CTDE paradigm, because the value function in QMix also serves as the policy in execution, and it will raise the inconsistent issue if you use the Feature-Pruned Agent-Specific Global States only in training and use the local observations in execution.
Is that right?

from on-policy.

Thanks for the reply! Another question: Does share_obs in line 34 in smac_runner.py represents FP states?

from on-policy.

I see. Once again thanks for your reply!

from on-policy.