LSTM Practical Exercise R. Q. Feitosa, J. D. Bermudez, J. A. Chamorro
Objective Predict the value of 𝑋 𝑡+1 from past observations ( 𝑋 𝑡−2 , 𝑋 𝑡−1 , 𝑋 𝑡 )
Network Architecture Architecture: Many-to-one LSTM
Dataset Daily USD to BRL exchange rate from April 30th 2018 to April 30th 2019.
Dataset Daily USD to BRL exchange rate from April 30th 2018 to April 30th 2019. Train Test
Dataset Dataset preprocessing: Sliding window 𝒙 𝟏 𝒙 𝟐 𝒙 𝟑 y 3.92 3.99 3.9 3.86 Train Test
Dataset Dataset preprocessing: Sliding window 𝒙 𝟏 𝒙 𝟐 𝒙 𝟑 y 3.92 3.99 3.9 3.86 3.89 Train Test
Dataset Dataset preprocessing: Sliding window 𝒙 𝟏 𝒙 𝟐 𝒙 𝟑 𝑦 3.92 3.99 3.9 3.86 3.89 … Train Test
Experimental setup Network architecture: LSTM (Many-to-one)
Expected Results Root Mean Squared Error Train: 0.02939 Test: 0.03424 Train prediction results Test prediction results
Exercises Exercise 1: Exercise 2: Define the LSTM network architecture according to this presentation Exercise 2: Evaluate the network performance in terms of RMSE and prediction graph for the next configuration modifications: 10 and 100 epochs 10 and 100 LSTM hidden units 2 LSTM layers (Deep LSTM) Use GRU instead of LSTM