I am excited to share our research paper, “Sequential Decision-Making in Atari 2600 Games: Comparing Temporal Features”.
This work is part of the thesis project of the talented Joao Martin–Saquet! It is the first step toward improving the use of computers by disabled persons. It highlights the best way to collect information about the user’s behaviour, so that we can assist them efficiently in their tasks.
Congratulations Joao for the outstanding work!
The paper shows how a temporal system can be described to efficiently apply reinforcement algorithms. We compared different temporal information and learning methods, such as LSTM, to determine the optimal parameters when solving a sequential task.
The results were surprising, as pre-processed and handcrafted features did not seem as good as initially thought. A new addition to the idea that neural networks perform better with raw information!
Here the link to the paper: https://ieeexplore.ieee.org/document/10974210
Ongoing work is being done to model this behaviour using genetic programming, with promising results!
Stay tuned and follow Joao Martin–Saquet if you are interested in more!