Improvement performance of IQN

[erinn-lee]

Please describe the feature you want to add.
A clear and concise description of what the feature. Ex. I'm going to implement ...

Improvement performance of IQN

Additional requirement
A clear and concise description of additional requirement for the new feature

Reference
Please append the reference about the feature

[erinn-lee]

Benchmarks of JORLDY agents
https://www.notion.so/Benchmark-09684f1adf764c84a5a331cb5690544f

Models with IQN networks have poor performance.
[ I ] Agents which series of IQN have lower performances than other Distributional RL agents.
[ II ] The n-step option tends to destabilize the performance of the Rainbow IQN

"Agents which series of IQN have lower performances than other Distributional RL agents"

IQN agent has lower or same performance comparing with C51 and QR-DQN. Please, refer the link on top.
M-IQN which applied 'Muchausen' RL technique also has lower or same performance too.
Their performances should be enhanced.

"The n-step option tends to destabilize the performance of the Rainbow IQN"

The performance of Rainbow IQN is unstable. Especially, it is vulnerable about Breakout task when the agent update by using n-step TD error. Their performances should be enhanced.