Zdrojaky:
narmaSRte.py
Mackey-Glass TE distribution
Parametre siete
TE reservoir distribution
Pozorovanie: Pri troche predstavivosti TE medzi jednotlivými neurónmi pripomína normálne rozdelenie. (teda v prípade, že TE je väčšie) A váhy rezervoárovej matice sú z normálenho rozdelenia.
TE reservoir distribution (instances of reservoir matrix)
TE reservoir-output distribution
Zdrojaky:
grafy.py
MGTEnormalized.py
TE a NARMA
TE a predikcie NARMA.
NRMSE
Mean TE
Mean TE reservoir-output
TE Input/Output
Zdrojaky:
NARMATE.py
NARMATEnormalized.py
NARMATEnormalizedOG.py
NARMATEnormalizedON.py
TE a Mackey-Glass
TE a predikcie Mackey-Glass.
NRMSE
Mean TE
Mean TE reservoir-output
TE Input/Output
Zdrojaky:
MGTE.py
MGTEnormalized.py
MGTEnormalizedOG.py
MGTEnormalizedON.py
TE a memory capacity
TE a úloha memory capacity. Input do siete bol Unif[-1,1].
Memory capacity
Mean TE
Mean TE reservoir-output
TE Input/Output
Zdrojaky:
TEMC.py
TEMCnormalized.py
TEMCnormalizedOG.py
TEMCnormalizedON.py
Overfitting
Pozeral som sa na overfitting ako na problém veľkých chýb a nakoniec bol problém v chybe v kóde.
Parametre siete
mean NRMSE test data
mean NRMSE train data
Zdrojaky:
standart.py
standartOverfit.py
normalized.py
normalizedOverfit.py
normalizedOG.py
normalizedOverfitOG.py
normalizedON.py
normalizedOverfitON.py
16: MG, NARMA & TE cont.
To isté čo pri #20 akurát s väčším rezervoárom + pridané korelácie medzi váhami rezervoárovej matice a transfer entropie tj. 
Parametre siete
| Standart | Normalized | Normalized & OG | Normalized & ON | |
|---|---|---|---|---|
| Veľkosť rezervoiru | 100 | 100 | 100 | 100 |
| Train Length | 1000 | 1000 | 1000 | 1000 |
| Test length | 1000 | 1000 | 1000 | 1000 |
| Runs | 10 | 10 | 10 | 10 |
| W | N(0,1) | N(0,1) | N(0,1) | N(0,1) |
| Win MG | Unif[-0.5,0.5] | Unif[-0.5,0.5] | Unif[-0.5,0.5] | Unif[-0.5,0.5] |
| Win NARMA | Unif[0,0.4] | Unif[0,0.4] | Unif[0,0.4] | Unif[0,0.4] |
| Spectral radius | unscaled | 0.9 | 0.9 | 0.9 |
| Ortogonal | null | null | gradient descent | gradient descent |
| Wout training | pinv | pinv | pinv | pinv |
| Bias | false | false | false | false |
| Leaking rate | 1 | 1 | 1 | 1 |
| OG/ON iterations | 0 | 0 | 30 | 30 |
| OG/ON eta | 0 | 0 | 0.03 | 0.07*0.9**t |
Parametre estimátorov
| k_history | k_tau | l_history | l_tau | u | k | |
|---|---|---|---|---|---|---|
| TE | 1 | 1 | 1 | 1 | 1 | 4 |
| AIS | 1 | 1 | null | null | null | 4 |
Mackey – Glass
Mean reservoir
| Standart | Normalized | Normalized & OG | Normalized & ON | |
|---|---|---|---|---|
| Mackey - Glass | 0.0096 | 0.6245 | 0.6117 | 0.6351 |
| NARMA | 0.0093 | 0.0922 | 0.0801 | 0.0945 |
NRMSE
| Standart | Normalized | Normalized & OG | Normalized & ON | |
|---|---|---|---|---|
| Mackey - Glass | 1.1149 | 0.9577 | 0.6748 | 1.0472 |
| NARMA | 2.4119 | 0.8664 | 0.8338 | 0.8433 |
InputOutput TE
| Standart | Normalized | Normalized & OG | Normalized & ON | |
|---|---|---|---|---|
| Mackey - Glass | -0.004 | -0.0210 | -0.0212 | -0.0256 |
| NARMA | -0.0013 | 0.3732 | 0.3743 | 0.3714 |
Mean reservoir-output TE
| Standart | Normalized | Normalized & OG | Normalized & ON | |
|---|---|---|---|---|
| Mackey - Glass | -0.0001 | 0.3516 | 0.3601 | 0.3532 |
| NARMA | -0.0006 | 0.0015 | 0.0014 | 0.0018 |
NARMA
AIS
Mean reservoir AIS
| Standart | Normalized | Normalized & OG | Normalized & ON | |
|---|---|---|---|---|
| Mackey - Glass | 0.0036 | 2.1694 | 2.1763 | 2.1910 |
| NARMA | 0.0023 | 0.1618 | 0.1675 | 0.1730 |
MG, NARMA TE
Prvé runny použitia ESNtoolbox na analýzu ESN trénovanej na úlohu predikcie t+1 Mackey – Glass a NARMA.
Parametre siete
| Standart | Normalized | Normalized & Orto | |
|---|---|---|---|
| Veľkosť rezervoiru | 10 | 10 | 10 |
| Train Length | 1000 | 1000 | 1000 |
| Test length | 1000 | 1000 | 1000 |
| Runs | 10 | 10 | 10 |
| W | Unif[0,1] | N(0,1) | N(0,1) |
| Win | Unif[0,1] | Unif[-0.1,0.1] | Unif[-0.1,0.1] |
| Spectral radius | unscaled | 0.9 | 0.9 |
| Ortonormal | null | null | QR |
| Wout training | pseudoinverse | pseudoinverse | pseudoinverse |
| Bias | false | false | false |
| Leaking rate | 1 | 1 | 1 |
Parametre estimátorov
| k_history | k_tau | l_history | l_tau | u | k | |
|---|---|---|---|---|---|---|
| TE | 1 | 1 | 1 | 1 | 1 | 4 |
| AIS | 1 | 1 | null | null | null | 4 |
NRMSE
| Standart | Normalized | Normalized & Orto | |
|---|---|---|---|
| Mackey - Glass | 0.035 | 0.217 | 0.111 |
| NARMA | 0.878 | 0.898 | 0.966 |
TE
Mean reservoir
| Standart | Normalized | Normalized & Orto | |
|---|---|---|---|
| Mackey - Glass | 0.006 | 0.590 | 0.322 |
| NARMA | -0.010 | 0.158 | 0.145 |
InputOutput TE
| Standart | Normalized | Normalized & Orto | |
|---|---|---|---|
| Mackey - Glass | 0.138 | -0.003 | 0.321 |
| NARMA | 1.811 | 0.867 | 0.266 |
AIS
Mean-reservoir AIS
| Standart | Normalized | Normalized & Orto | |
|---|---|---|---|
| Mackey - Glass | 2.277 | 2.230 | 1.687 |
| NARMA | 0.392 | 0.354 | 0.110 |
Zdrojaky:
PS
Markov models from data by simple nonlinear time series predictors in delay embedding spaces
ESN Toolbox
Dal som dokopy knižnicu, ktorá zatiaľ implementuje:
KSG Transfer entropy estimator – TE (source, target, kHistory, kTau, lHistory, lTau, u, k)
KSG Active information storage estimator – AIS (target, kHistory, kTau, k)
ESN reservoir Memory capacity – MC(Win, W)
Zdrojaky:
ESNtoolbox.py
