Use different lr for different policies

[call-me-anything-you]

I’m using rllib under the MARL setting and need to set different lr for different policies. I’ve tried to set these in configs.multi_agent() as follows:

configs.multi_agent(
    policies = {
        "policy_0": (None, None, None, {"lr": 1e-4}), 
        "policy_1": (None, None, None, {"lr": 2e-4}), 
    }
)

However, it seems like none of these works. The policies still use the lr set in

configs.training(lr=xxx)

How can I do this right?