machine learning - Prediction of following sequence based on small sequence -

i totally new in machine learning , have problem, want solve using ai or so. appreciate, if recommend me concrete algorithms, neural network architectures or related reading.

i doing research predicting users intent based on mouse movement. in phase of analysis without concrete dataset. goal predict target of users's intention (e.g. button, user click) predicting mouse trajectory.

let me introduce problem

i have lot of sequences. length of each sequence may vary. input pass smaller sequence, want predict next x values. want know next possible sequence (or more possible sequences). length of output sequence (x) variable. maybe sequence ends here? prediction should done in “real time”.

so sequences?

sequence represents directions of movement in 2-dimensional space after preprocessing. each value integer of interval <0,8>. algorithm should capable of increasing upper limit of interval (16, 32, ...). actually, value interpolated angle.

three example sequences. real sequences bigger.

how imagine solution?

sequences clustered based on similarities. when dataset of sequences made, neural network trained retrieve sequences, contain input sequence subsequence, quick possible.

clustering

matching subsequence should have tolerance. sequence [3, 3, 3, 3, 2] similar [3, 3, 4, 3, 2] = deviation tolerance*. or sequence [4, 3, 3, 2] similar [4, 3, 3, 3, 3, 2] = tolerance on values repeated continuously.

*i can tell difference between 2 values relative number - 0% same direction => 100% opposite direction.

if input [ 1,2,2,2 ] - red - output should [ 4,3,2,2 ].

if input [ 3,3,3,2 ] - blue - output should [ 2 ].

neural network

after research found hopfield network, should give similar sequence. realized sequences lengths variable , hopfield network architecture expects binary values.

i somehow create binary representation of sequence have no idea how manage lengths may vary.

let’s make level

what if every value in sequence not scalar velocity vector (d, s), d direction , s speed?

related questions

1. can neural networks trained “online”? no need know previous train dataset, give new dataset.
2. can neural networks trained on server side (e.g. python) used prediction on client side (javascript)?
3. can neural networks have kind of “short term memory” - prediction affected 2-3 previous predictions?
4. most important - should use neural networks or approach?

thanks everyone.

feel free correct english.

can neural networks trained “online”? no need know previous train dataset, give new dataset.

typically, dont train ann continuously. train until error within tolerance, use model going forward make predictions. store information off , retrain network every night if want periodically adjust model, odds that's not going offer improvement, , runs risk of prolonged bad data of skewing model.

can neural networks trained on server side (e.g. python) used prediction on client side (javascript)?

it depends. intend use trained model client prediction, or intend user actions live-train model used prediction? if model trained, can use prediction of user events. if model not trained, run risk of bad data corrupting model. live-training require constant update of model settings on client side new model generated server.

can neural networks have kind of “short term memory” - prediction affected 2-3 previous predictions?

using previous predictions input not recommended. introduces entropy system can allow model drastically deviate reliable predictions if makes few bad predictions in row. can try it, in case you'll need n*k additional nodes on input layer, n number of previous predictions want use, , k number of output values in prediction.

most important - should use neural networks or approach?

anns useful predicting things. biggest problem defining scope, , relevant reliable data necessary make prediction. i've made anns predict market volatility in video games, thousands of input values, predicting mouse movements going challenge. nothing stopping user moving mouse in circle hours, or leaving cursor in 1 spot. each time sample such action, going make model more predict type of behavior. training data, , controlled environment essential. video games make bad environment predicting mouse movement, user behavior dependent on more previous mouse movements. websites favorable environment though, during session, user navigates in predictable ways through finite space.