15 lines
1.2 KiB
TeX
15 lines
1.2 KiB
TeX
|
\section{Background}
|
||
|
|
% what other people did that is closely related to yours.
|
||
|
|
|
||
|
|
Our work is heavily based off the research done by Minh et. al in \textit{Human-Level Control through Deep Reinforcement Learning} \cite{humanlevel}. The algorithm we developed to solve \textit{Celeste Classic} uses a deep Q-learning algorithm supported by replay memory, with a modified reward system and explore-exploit probability. This is very similar to the architecture presented by Minh et al.
|
||
|
|
|
||
|
|
\vspace{2mm}
|
||
|
|
|
||
|
|
The greatest difference between our approach and the approach of \textit{Human-Level Control} is the input space and neural network type. Minh et. al use a convolutional neural network, which takes the game screen as input. This requires a significant amount of training epochs and computation time, and was thus an unreasonable approach for us. We instead used a plain linear neural network with two inputs: player x and player y.
|
||
|
|
|
||
|
|
\vspace{2mm}
|
||
|
|
|
||
|
|
Another project similar to ours is AiSpawn's \textit{AI Learns to Speedrun Celeste} \cite{aispawn}. Here, AiSpawn completes the same task we do---solving \textit{Celeste Classic}---but he uses a completely different, evolution-based approach.
|
||
|
|
|
||
|
|
\vfill
|
||
|
|
\pagebreak
|