Need for Prediction How Can We Predict: . . . Examples of Models How Do We Estimate . . . Efficient Parameter-Estimating Often, the Empirical . . . Algorithms for Symmetry-Motivated Computationally . . . Models: Econometrics and Beyond How Can We Easily . . . Shift-Invariant Case: . . . Vladik Kreinovich 1 , Anh H. Ly 2 , Olga Kosheleva 1 Scale-Invariant Case: . . . and Songsak Sriboonchitta 3 Home Page 1 University of Texas at El Paso, USA Title Page olgak@utep.edu, vladik@utep.edu 2 Banking University of Ho Chi Minh City, 56 Hoang Dieu 2 ◭◭ ◮◮ Quan Thu Duc, Thu Duc, Ho Ch´ ı Minh City 3 Faculty of Economics, Chiang Mai University ◭ ◮ Chiang Mai 50200 Thailand, songsak econ@gmail.com Page 1 of 39 Go Back Full Screen Close Quit
Need for Prediction How Can We Predict: . . . 1. Need for Prediction Examples of Models • In many real-life situations, we have a quantity x that How Do We Estimate . . . changes with time t . Often, the Empirical . . . Computationally . . . • We want to use the previous values of this quantity to How Can We Easily . . . predict its future values. Shift-Invariant Case: . . . • For example: Scale-Invariant Case: . . . – we know how the stock price has changed with time, Home Page and Title Page – we want to use this information to predict future ◭◭ ◮◮ stock prices. ◭ ◮ • In many cases, such a prediction is possible; for exam- Page 2 of 39 ple: Go Back – when weather records show clear yearly cycles, Full Screen – it is reasonable to predict that a similar yearly cycle will be observed in the future as well. Close Quit
Need for Prediction How Can We Predict: . . . 2. How Can We Predict: Main Idea Examples of Models • A usual approach to prediction is that we select some How Do We Estimate . . . model , i.e., some parametric family of functions Often, the Empirical . . . Computationally . . . f ( t, c 1 , . . . , c ℓ ) . How Can We Easily . . . • Based on the available observations, we find the pa- Shift-Invariant Case: . . . rameters � c i which provide the best fit. Scale-Invariant Case: . . . Home Page • hen we use these values � c j to predict the future values Title Page of the quantity x as x ( t ) ≈ f ( t, � c 1 , . . . , � c ℓ ) . ◭◭ ◮◮ ◭ ◮ Page 3 of 39 Go Back Full Screen Close Quit
Need for Prediction How Can We Predict: . . . 3. Examples of Models Examples of Models • In some cases, the dependence of the quantity x on How Do We Estimate . . . time t is polynomial, in which case Often, the Empirical . . . f ( t, c 1 , . . . , c ℓ ) = c 1 + c 2 · t + c 3 · t 2 + . . . + c ℓ · t ℓ − 1 . Computationally . . . How Can We Easily . . . • For a simple periodic process, the dependence of the Shift-Invariant Case: . . . quantity x on time is described by a sinusoid: Scale-Invariant Case: . . . Home Page f ( t, c 1 , c 2 , c 3 ) = c 1 · sin( c 2 · t + c 3 ) . Title Page • To get a more realistic description of a periodic process, ◭◭ ◮◮ we need to take into account higher harmonics: ◭ ◮ f ( t, c 1 , c 2 , . . . ) = c 1 · sin( c 2 · t + c 3 )+ c 4 · sin(2 c 2 · t + c 5 )+ . . . Page 4 of 39 • For a simple radioactive decay, the amount of radioac- Go Back tive material decreases exponentially: Full Screen f ( t, c 1 , c 2 ) = c 1 · exp( − c 2 · t ) . Close Quit
Need for Prediction How Can We Predict: . . . 4. Examples of Models (cont-d) Examples of Models • A more realistic model is a mixture of several different How Do We Estimate . . . isotopes, with different half-lives: Often, the Empirical . . . Computationally . . . f ( t, c 1 , c 2 , . . . ) = c 1 · exp( − c 2 · t ) + c 3 · exp( − c 4 · t ) + . . . How Can We Easily . . . Shift-Invariant Case: . . . • Other models include log-periodic model which is used to predict economic crashes: Scale-Invariant Case: . . . Home Page c 1 + c 2 · ( c 3 − t ) c 4 + c 5 · ( c 3 − t ) c 4 · cos( c 6 · ln( c 3 − t ) + c 7 ) . Title Page • The following software model describes the number of ◭◭ ◮◮ bugs discovered by time t : ◭ ◮ f ( t, c 1 , c 2 , c 3 ) = c 1 · ln( t − c 2 ) + c 3 . Page 5 of 39 Go Back • A more complex example is a neural network, when c j are the corresponding weights. Full Screen Close Quit
Need for Prediction How Can We Predict: . . . 5. How Do We Estimate the Parameters? Examples of Models • Usually, the Least Squares method is used to estimate How Do We Estimate . . . the values of the parameters c 1 , . . . , c ℓ . Often, the Empirical . . . Computationally . . . • So, based on the observed values x ( t i ), we find c j that � n How Can We Easily . . . ( x i − f ( t i , c 1 , . . . , c ℓ )) 2 . minimize Shift-Invariant Case: . . . i =1 Scale-Invariant Case: . . . • In some cases – e.g., for the polynomial dependence – Home Page the model f ( x, c 1 , . . . , c ℓ ) linearly depends on c j . Title Page • Then, the minimized expression is quadratic in c j . ◭◭ ◮◮ • We can find the minimum of a function of several vari- ◭ ◮ ables by equating all its partial derivatives to 0. Page 6 of 39 • For a quadratic objective function, all the partial Go Back derivatives are linear functions of c j . Full Screen Close Quit
Need for Prediction How Can We Predict: . . . 6. How Do We Estimate the Parameters (cont-d) Examples of Models • Thus, by equating them all to 0, we get a system of How Do We Estimate . . . linear equations for the unknowns c j . Often, the Empirical . . . Computationally . . . • For solving systems of linear equations, there are many How Can We Easily . . . efficient algorithms. Shift-Invariant Case: . . . • So in this case, the problem of identifying the model’s Scale-Invariant Case: . . . parameters is computationally easy. Home Page • On the other hand, in general, the dependence of the Title Page model on the parameters c j is non-linear. ◭◭ ◮◮ • Thus, the objective function is more complex than ◭ ◮ quadratic. Page 7 of 39 • It is known that, in general, optimization is computa- Go Back tionally intensive – NP-hard. Full Screen • It is therefore desirable to select models for which iden- tification is easier. But how do we select modles? Close Quit
Need for Prediction How Can We Predict: . . . 7. How Are Models Selected in the First Place? Examples of Models • Sometimes, we have an good understanding of the pro- How Do We Estimate . . . cesses that cause the quantity x to change. Often, the Empirical . . . Computationally . . . • In such situations, we have a theoretically justified How Can We Easily . . . model. Shift-Invariant Case: . . . • In most cases, however, the model is selected empiri- Scale-Invariant Case: . . . cally: Home Page – we try different models, and Title Page – we select the one for which, for the same number ◭◭ ◮◮ of parameters, the approximation error is min. ◭ ◮ Page 8 of 39 Go Back Full Screen Close Quit
Need for Prediction How Can We Predict: . . . 8. Often, the Empirical Efficiency of Selected Examples of Models Models Can Be Explained by Symmetry How Do We Estimate . . . • In an empirical choice, we only compare a few possible Often, the Empirical . . . models. Computationally . . . How Can We Easily . . . • As a result Shift-Invariant Case: . . . – the fact that the selected model turned out to be Scale-Invariant Case: . . . better than others Home Page – does not necessarily mean that this model is indeed Title Page the best for a given phenomenon: ◭◭ ◮◮ – there are, in principle, many other models that we ◭ ◮ did not consider in our empirical comparison. Page 9 of 39 • Good news is that in many cases, the empirical selec- tion can be confirmed by a theoretical analysis. Go Back • Often, the empirically successful model can be derived Full Screen from the natural symmetry requirements. Close Quit
Need for Prediction How Can We Predict: . . . 9. But the Model Remains Computationally In- Examples of Models tensive How Do We Estimate . . . • The fact that the empirically selected model is theo- Often, the Empirical . . . retically justified does not change its formulas; so: Computationally . . . How Can We Easily . . . – if the dependence of this model on the correspond- Shift-Invariant Case: . . . ing parameters c j is non-linear, Scale-Invariant Case: . . . – the problem of identifying parameters of this model Home Page remains computationally intensive. Title Page • In this talk, we show that symmetries: ◭◭ ◮◮ – are not only helpful in selecting a model, ◭ ◮ – they can also help design computationally efficient Page 10 of 39 algorithms for identifying model’s parameters. Go Back Full Screen Close Quit
Recommend
More recommend
