search

Guiliang / Sport-Analytic-NN

Neural Network Realization of Project Sport Analytic
3 stars 1 forks source link

RNN(LSTM) Network does not converge #1

Open Guiliang opened 7 years ago

Guiliang commented 7 years ago

Solving issue 1

  1. Add a fully connected dense layer over the output of network.
  2. [fail] Coordinate some parameters like "keep-Probability" or "Number of LSTM Layers".
  3. Try to balance the data, for example number of 0 and 1 of target y
  4. Coordinate the x and y axis of network
Guiliang commented 7 years ago
  1. Try to focus on the difference of V between two team (set reward to goal difference?)
  2. Try batch Normalization
Guiliang commented 7 years ago

It seems "Coordinate parameters" doesn't work.

Guiliang commented 7 years ago
  1. [Work] reduce the length of trace data (From 10 to 3)
Guiliang commented 7 years ago

[Work] reduce the length of trace data (From 10 to 3) does work, but it will reduce the trace power of LSTM

Guiliang commented 7 years ago
  1. Try to initialize the weight of LSTM with smaller numbers (uniformly in [-0.08, 0.08]).
  2. Try larger hidden size and layer size.