Small Steps: A Experimental Case for Compound Prediction

These results would not be suitable for any engineering task and are suboptimal.Now let’s look at the results of the compound prediction method throughout the prediction progression.Predicted values of theta based on link lengths and input Theta 2First order kinematic coefficients ( h ) predictions based on link lengths, input theta 2, and theta prediction valuesAngular velocities ( w ) based on link lengths, predicted theta values and first order coefficientsSecond order kinematic coefficients ( h’ ) based on link lengths, predicted theta values, first order coefficients, and angular velocitiesDiscussionBased on the coefficients of determination, the results of the “compound prediction” method were far more accurate than the “giant leap” method (an average R² of 0.9773 vs 0.8745)..This is true despite the compounding error that resulted from making predictions based on predicted values..Given the necessary input parameters, the relationship between the input and output variables is simply too abstract for the current model used in the “giant leap” method..On the other hand, the model is sufficiently robust to make accurate sequential predictions of intermediate variables eventually reaching an accurate prediction of the final output.There are several limitations to these results that I would like to acknowledge; although, this is far from a complete list..First, the size of the network was not changed between the cases..Perhaps if enough tuning was done to the network used for the “giant leap” method that it would have better accuracy..Second, the test size was only a few thousand data points, having more may have improved the results..Third, the network parameters were chosen by a few trial-and-error iterations and are in no way intended to be understood as the optimal network for this particular task..Fourth, the coefficient of determination may not be the best metric on which to base the effectiveness of a given model.Conclusion/TLDRWhat I hope to convey through this article and small experiment is that there may be cases in which making smaller, simpler, predictions leading up to the final prediction is a strategy worth exploring!.For example, instead of predicting the value of a home based on a pile of metrics, predict the value of the land, the value of the structure, the value of the furniture/appliances, etc..and then make the final prediction based on those values.I in no way intend to convey the results of this test as the general rule, simply as an thought-provoking example to stage the question for further, more detailed, exploration.Please reach out with any questions..Also, if anyone would like the dataset, please message me and I can make it available.. More details

Leave a Reply