"Decision Transformer vs. Decision Mamba: Analysing the Complexity of Sequential Decision Making in Atari Games"

Comparing Decision Transformer vs Decision Mamba reveals each model's sweet spot in game complexity.

This paper analyzes performance differences between Decision Transformer (DT) and Decision Mamba (DM) across 12 Atari games, revealing that action space complexity and visual complexity significantly influence their relative performance. DT excels in complex environments while DM performs better in simpler ones.

-----

https://arxiv.org/abs/2412.00725

🎮 Original Problem:

→ No comprehensive analysis exists comparing Decision Transformer and Decision Mamba's performance across different game environments, leaving uncertainty about which model performs better in specific scenarios.

-----

🔍 Solution in this Paper:

→ The study analyzed 12 Atari games using metrics like action space complexity, visual complexity, and trajectory length.

→ Implemented visual complexity quantification using image entropy, compression ratio, and feature count.

→ Applied random forest regression and correlation analysis to identify key performance factors.

→ Developed action fusion strategies to isolate action space complexity effects.

-----

🎯 Key Insights:

→ Action space complexity is the primary factor affecting model performance

→ Visual complexity, measured by compression ratio, significantly influences model effectiveness

→ DM excels in visually simpler environments with fewer actions

→ DT shows advantages in games with higher visual and action complexity

-----

📊 Results:

→ DT outperformed in 8 of 12 games, including Hero and KungFuMaster

→ DM showed superior performance in Breakout with normalized scores of 401.39 vs DT's 238.19

→ Action fusion reduced DT's performance in Hero from 30.37 to 18.72, while DM remained stable