Sequential Decision-Making in Atari 2600 Games: Comparing Temporal Features

This paper highlights the importance of incorporating temporal information into features to effectively address an environment derived from an Atari 2600 game that requires time handling. We present a study evaluating which features or models are better suited for solving tasks involving sequential decisions. Our approach includes comparing several methods, such as using temporal features (handcrafted features or sequences of environment states) or an LSTM, focusing on their ability to capture temporal dependencies. This study presents notable results on convergence speeds and model stability, highlighting performances on features and models used, such as handcrafted features, sequence of states, and Long-Short Term Memory model.