Dropping out agents while training on rl examples

[CriZ-sd]

Hello,

i noticed that while training in rl examples, some of the agents are dropped out. I have also two images one from a custom example (with three agents) and one from the cartpole env example that is included in the ibmfl examples (with five agents).

As it is depicted in both cases, the agents are gradually dropped out and finally, only one agent remains at the training processes. Is this the expected behavior ? Any idea why this happens?