[Question] What is the point of having a separate .rewards dictionary?

[daxmavy]

Question

As far as I can tell, rewards are managed in the following manner:

1. _cumulative_rewards[agent] is returned by env.last() (along with observation, termination, truncation, info)
2. Policy chooses an action, which is then executed by env.step(action).

I know that rewards are more complicated than in the typical RL environment, because the reward for agent_0 should in some circumstances be adjusted during the turn of agent_1, for example. IIUC, _cumulative_rewards is used to account for this.

Within step (in 2. above), the following occurs:

1. We set self._cumulative_rewards[agent] = 0, because the policy for agent has already received this reward from env.last() and processed it while choosing a next action.
2. The dictionary self.rewards is updated according to the consequences of action.
3. We then use self._accumulate_rewards() to update the self._cumulative_rewards dictionary. From this code, this straightforwardly just increments self._cumulative_rewards with the values in self.rewards.

I have two questions:

1. Why is self.rewards even needed? Why not just directly adjust self._cumulative_rewards to incorporate the consequences of action?
2. Am I correct in thinking that self.rewards should be set to 0 for all agents after every call to self._accumulate_rewards()? My reasoning is that otherwise, the values in rewards will be added multiple times to _cumulative_rewards, which is undesirable. If this is the case, why isn't this functionality built into self._cumulative_rewards?