Prism is a modular system of deep Q-learning algorithms and modifications that can be combined.
IQN
IDS
DQN
PER
Double Q Learning
Layer Normalization
N-Step Returns
The default configuration is IDS + IQN + Layer Normalization + 3-step returns with a target network. This seemed to perform the best across the MinAtar experiments.
To find the best combination of the components available, I ran additive and subtractive ablation studies similar to the Revisiting Rainbow paper. Data from the studies are available on WandB. The additive study is here and the subtractive study is here. Due to time constraints, I only ran experiments in MinAtar. Hyperparameters can be found in the corresponding config and experiment files in this repo.