Continuous And Discrete Action Space Handling With Unified Latent Planning

via “gameaction-discrete-action-space”

Abstract reasoning benchmark with $1M prize for AGI.

Unique: Abstracts task interaction through a discrete GameAction enum, providing a consistent interface across all benchmark tasks. Action semantics are abstracted, enabling agents to learn action effects through observation rather than explicit specification.

vs others: More standardized than task-specific action interfaces by providing a unified enum; more flexible than fixed action sets by supporting task-agnostic action selection.

* ⏫ 02/2023: [Grounding Large Language Models in Interactive Environments with Online RL (GLAM)](https://arxiv.org/abs/2302.02662)

Unique: DreamerV3 uses a single latent-space policy architecture that parameterizes both continuous and discrete action distributions without task-specific modifications, treating action space type as a hyperparameter rather than an architectural choice. This contrasts with prior work that required separate policy heads or explicit action space handling.

vs others: Enables unified training across Atari and continuous control benchmarks with identical code, whereas most RL frameworks require separate implementations or significant hyperparameter tuning per domain.