I am trying to make a model with LSTM for chess move prediction.
I have a dataset of games and for each game I have the list of moves.
#for example:
[b3, c5, Bb2, Nc6, g3, d6, Bg2, Nf6, c4, a6, Nc3, ...]
Since LSTM models don't process stings I converted the moves with a keras tokenizer. More specifically I first made a set of moves and then encoded them as integers. Once this procedure was done I put these integers in a list like this :
[7464, 2678, 2467, 1114, 6944, 2801, 7937, 7743, 1503, 6348, 5215, ...]
from this list I created two other lists called X and Y. These other two lists are formed like this:
#X:
[[7464, 2678, 2467, 1114, 6944],
[2678, 2467, 1114, 6944, 2801],
[2467, 1114, 6944, 2801, 7937],
[1114, 6944, 2801, 7937, 7743],
[6944, 2801, 7937, 7743, 1503],
[2801, 7937, 7743, 1503, 6348],
[7937, 7743, 1503, 6348, 5215], ...]
#Y:
[2801, 7937, 7743, 1503, 6348, 5215, ...]
Once I created X and Y I converted them into a numpy array and reshaped X as follows:
n_features = 1
X = X.reshape((X.shape[0], X.shape[1], n_features))
#X:
[[[7464]
[2678]
[2467]
[1114]
[6944]]
[[2678]
[2467]
[1114]
[6944]
[2801]] ...]
#Y:
[2801 7937 7743 1503 6348 5215 ...]
Before moving on to the model I have to make a small premise.
Since this is a first run I passed the model to fit just one batch to see if everything works well.
Then I passed these two structures modeled as follows to the model below to do the FIT:
X= np.asarray(X).astype(np.int)
Y= np.asarray(Y).astype(np.int)
model = tf.keras.Sequential()
model.add(layers.LSTM(50, activation='relu', input_shape=(n_steps, n_features)))
model.add(layers.Dense(1))
model.compile(optimizer=tf.keras.optimizers.Adam(0.01), loss=tf.keras.losses.MeanSquaredError(), metrics=['accuracy'])
model.fit(X, Y, epochs=200, verbose=1)
and then I try to make a prediction with the first 5 elements of the match that I passed to it as a train but the prediction is wrong. In addition to returning a move that is not possible to make, it returns a different move from the one it learned and since I did the train with only one match, I think it should give me a known result.
#A code example for predicting:
test_data = np.array([7464, 2678, 2467 ,1114 ,6944])
test_data = test_data.reshape((1, n_steps, n_features))
predictNextNumber = model.predict(test_data, verbose=1)
print(predictNextNumber)
predictNextNumber= int(predictNextNumber)
text = tokenizer.sequences_to_texts([[predictNextNumber]])
print(text)
#Output:
1/1 [==============================] - 0s 493ms/step
[[1061.2649]]
['g1=q#']
Can anyone tell me where I am going wrong?
I have recently considered an option. Is it possible that the model I have been using is too much based on integer prediction and therefore looking for a numerical relationship between the input and what it has learned? Because in this case I would need a model that still learns the positions of the numbers and not a mathematical correlation between them.
One last question, once everything works with a match, how do I pass at least half of the matches in the dataset to the model to make in more consistent fit?
