Real World Examples of Agent based Decision Support Systems for Deep

Learning based on Complex Feed Forward Neural Networks

Harald R. Kisch and Claudia L. R. Motta

NCE, Universidade Federal do Rio de Janeiro, Av. Pedro Calmon, 550 - Cidade Universitria, Rio de Janeiro, Brazil

Keywords:

Deep Learning, Neural Network, Feed Forward, Muilti-agent, Complex Systems, Decision Support.

Abstract:

Nature frequently shows us phenomena that in many cases are not fully understood. To research these phe-

nomena we use approaches in computer simulations. This article presents a model based approach for the

simulation of human brain functions in order to create recurrent machine learning map fractals that enable

the investigation of any problem trained beforehand. On top of a neural network for which each neuron is

illustrated with biological capabilities like collection, association, operation, deﬁnition and transformation,

a thinking model for imagination and reasoning is exempliﬁed in this research. This research illustrates the

technical complexity of our dual thinking process in a mathematical and computational way and describes

two examples, where an adaptive and self-regulating learning process was applied to real world examples.

In conclusion, this research exempliﬁes how a previously researched conceptual model (SLA process) can

be used for making progress to simulate the complex systematics of human thinking processes and gives an

overview of the next major steps for making progress on how artiﬁcial intelligence can be used to simulate

natural learning.

1 INTRODUCTION

Many of our technical innovations are inspired by na-

ture. Most of the things we categorize in biologi-

cal evolution, like how human brains works, are still

difﬁcult to achieve with our current technical skills.

Backﬁring introduced by Hebb, 1949 with ”Firing to-

gether, wired together” and nowadays called back-

propagation or Forward- or Reverse-mode differen-

tiation (Ochs et al., 2016), which means reducing

many neural network paths by ﬁnding and propa-

gating intermediate results that signiﬁcantly reduce

path-reassemblings or different or additional path-

assimiliations (Krizhevsky et al., 2012). Recently, the

Forward-mode differentiation (tracking how one in-

put affects every node) and the Reverse-mode differ-

entiation (tracking how every node affects one out-

put) to mathematically evaluate derivatives were pre-

sented (Ochs et al., 2016). As an advantage of imple-

menting Reverse-mode differentiation it is possible to

get all the derivatives that produce the output network

streams with one single computation. This reduces

the computational costs of so called ﬁrst-order opti-

mization algorithms within learning functions e.g. by

ﬁnding a local minimum or maximum. The other way

around, implementing Forward-mode differentiation

means getting all derivatives that produce the input

with one single computation. Thus, we can choose

the computation depending on which operation per-

forms better with respect to the input or output se-

quences. If there is a computation with a lot of out-

puts, the chain rule is to use Forward-mode differen-

tiation to create deep feedforeward neural networks.

If there is a computation with a lot of input, the chain

rule is to use Reverse-mode differentiation to create

neural networks, whose derivatives are ﬂowing back-

wards through the model. The replacement of the lo-

gistic or the hidden networks is represented through

an analog value, which has its own input and output

gates that control when inputs are allowed to inﬂu-

ence the output. These gates have their own learned

weights on connections and memories that are stored

in the previous step. In the past, it seemed unlikely

that an exact algorithm can be developed to perform

probabilistic inference efﬁciently over all classes of

belief networks (Cooper, 1990). However, 20 years

later the theory of discrete calculus was introduced

(Grady and Polimeni, 2010), showing that it can be

applied in many ﬁelds and disciplines by unifying

approaches for data analysis and content extraction.

These extractions can speed up the optimization of

neural networks by the amount of input or output se-

Kisch, H. and Motta, C.

Real World Examples of Agent based Decision Support Systems for Deep Learning based on Complex Feed Forward Neural Networks.

DOI: 10.5220/0006307000940101

In Proceedings of the 2nd International Conference on Complexity, Future Information Systems and Risk (COMPLEXIS 2017), pages 94-101

ISBN: 978-989-758-244-8

quences in recurrent neural network (RNN) streams.

This leads to a faster learning of long term depen-

dencies in Long Short Term Memory (LSTM) Re-

current Neural Networks (Le et al., 2015). Hinton

et. al (2015) also exempliﬁes the solution of some

challenging problems like sequence regression, pre-

sentation and performance increase on hidden layer

establishment at initialization time. Some examples

of this approach can also be found in handwriting

recognition (Graves et al., 2009), handwriting genera-

tion (Graves, 2013), speech recognition (Graves et al.,

2013), (Graves and Jaitly, 2014), machine translation

(Luong et al., 2014), (Bahdanau et al., 2014) and se-

quence to sequence mapping (Sutskever et al., 2014).

In this article, we use the theory of Dual Thinking

(Scheffer et al., 2015) as a meta model and combine

it with a previous work on a human thinking model

based on a neural network called the 6-Step Learn-

ing Adaption process (SLA) (Kisch and Motta, 2015).

The main idea of SLA is the step-wise, agent ori-

ented approach to collect, associate, operate, deﬁne,

store and transfom knowledge map fractals in a tool-

validated conceptual model.

2 METHODOLOGY

Many scientists consistently conclude that intuition

and reasoning is an essential part of our thinking pro-

cess (Scheffer et al., 2015). In relation to the SLA

process intuition can be realized by mapping gate-

associations, where gates are deﬁnitions of neuronal

sub-networks according to Sporns motifs (Sporns and

Koetter, 2004). On the other hand reasoning in

the SLA-Process are the operations on the deﬁni-

tions of neural sub-networks which lead to deﬁnition-

increments into the overall learning map. Each map

fractal can span out the whole network from which it

was deﬁned. Each node in the network saves the state

of any other node related to the deﬁnition. This means

the collection at the beginning, the associations, the

operations that were made and also the transformation

including the previously deﬁned results and predic-

tions for the changes of the next transformation based

on the positive and negative feedback out of the prob-

ability testing network in a whole.

Notes: P = Product (Reasoning), T = Testing (Intu-

ition), nF = negative Feedback, pF = positive Feed-

back, G

= Testing-Gate, G

= Product Gate, D

∗

Current Deﬁnition

With both processes for reasoning and testing com-

bined, there are some feedback-networks with posi-

tive and negative feedback on the actual thinking pro-

cess called negative Feedback (nF) and positive Feed-

Figure 1: Dual Thinking.

Figure 2: Dual Thinking Layer.

back (pF) as shown in Figure 1 in the center of the

top and of the bottom circle, respectively. Negative

feedback (nF) happens primarily in the Intuition-Test

Neuronal Network (T ) and pF happens primarily in

the Reasoning-Neural Network (P) as shown in Fig-

ure 1 on the top circle and on the bottom circle. Feed-

back is the deﬁnition of a previous match of associ-

ations and can be positive (many matches) or nega-

tive (no matches). The feedback of many neural net-

works in this approach is generally stored as result

and used to reach group consensus according to Liao

and Ji (Liao and Ji, 2014).

3 RESULT

Many approaches, which are used to mathematically

describe group consensus (Ji et al., 2014) in neural

networks need many restrictions because they cannot

be applied directly to real world szenarios. This is

why the formalization approach in this article is a gen-

eral overview to explain the conceptual layer model

of ﬁgure 1. In the following, the formalization of the

Dual Thinking process illustrated in Figure 1 and 2

is conceptional formalized in general and meant to be

used in big data computation scenarios.

Real World Examples of Agent based Decision Support Systems for Deep Learning based on Complex Feed Forward Neural Networks

Implementation Setup Formalization.

AT T := JSON ∈ {Array, Function, Ob ject} (1)

Spec := Ob ject ∈ {key : value} (2)

DOC :=

∞

[

n=1

AT T + Spec + Rev# (3)

DB :=

∞

[

n=1

DOC (4)

N :=

∞

[

n=1

DB (5)

Note: ATT = Attachment, Spec = Speciﬁcation, DOC

= Document, DB = Database, N = Node, Rev# =

unique revision number, which is used to build a

concurrent versioning of any knowledge map frac-

tal. This is mandatory for the transformation step in

the SLA process to extract resonances of positive and

negative feedback (explained later on) without having

to change the knowledge at all. On negative feedback

for example, the state of the best feedback resonance

will be recovered using the saved Rev#.

Model Formalization.

NN :=

∞

[

n=1

N (6)

M :=

∞

[

n=1

NN (7)

G :=

∞

[

n=1

M (8)

D :=

∞

[

n=1

G (9)

S :=

∞

[

n=1

D (10)

Note: NN = Neural Network, M = Motif, G = Gate,

D = Deﬁnition, S = Stack item

The stack S is saved in every node N to span out the

whole NN to further recure the inputs leading to the

deﬁnition D. We can say that a reverse function which

returns all the inputs I of a neural network is deﬁned

as following:

I = f (S

) → NN

∪ D

(11)

The other way around, the reverse function D

∗

is the

weighted conjunction between G

∗

and G

∗

producing

the Output O and can deﬁned as:

O = f (G

∪ G

) → D

∗

(12)

Any NN consist of a testing network T which repre-

sents the imagination part of our thoughts and a pro-

ducing network P which represents the reasoning part

of our thoughts. The conclusions (deﬁnitions D ) of T

and P together are saved in hidden feedback networks

nF and pF .

nF = f (NN

) → O(G

);NN

, G

∈ D

∗

(13)

pF = f (NN

) → O(G

);NN

, G

∈ D

∗

(14)

Figure 1 illustrates (11) to (14). In this illustration,

each circle is meant to be a neural network or a part

of a neural network Motif (Sporns and Koetter, 2004)

itself. The Sequences S1 to S4 executes the Sequence

functions s

and s

as following:

S1 := f (T ) ∈ s

(15)

S2 := f (P) ∈ s

(16)

S3 := f (G

) ∈ s

(17)

S4 := f (G

) ∈ s

(18)

Sequences S1 to S4 in ﬁgure 1 are followed by se-

quences s

and s

, which can be deﬁned as following:

:= f (T ) ∪ f (P); T, P ∈ D

∗

(19)

:= f (G

) ∪ f (G

);G

, G

∈ D

∗

(20)

The feedback functions nF and pF are deﬁned as fol-

lowing where nF is meant to be the negative part and

pF the positive part of the feedback:

nF := f (T ) ∪ f (G

);T, G

∈ D

∗

(21)

pF := f (P) ∪ f (G

);P, G

∈ D

∗

(22)

The feedback functions conclude, that the following

deﬁnition D

is a neural network NN of positive in-

puts I

out of the current network M

∗

transformed

by the test network T , producing products P of pos-

itive feedback gates nF(G

) in disjunction with all

negative feedback gates nF(G

), so that the resulting

function for the next deﬁnition can be used as follow-

ing:

:= M

∗

∏



pF(G

) ∩ nF(G

)



∈ T ( f (S

))

(23)

COMPLEXIS 2017 - 2nd International Conference on Complexity, Future Information Systems and Risk

In most cases a complex problem needs an appro-

priate environment, in which the problem can be

solved. This environment needs training on the spec-

iﬁed inputs as well as a ﬁnishing condition with de-

ﬁned dropouts to prevent overﬁtting (Srivastava et al.,

2014).

4 DIFFERENCE TO TYPICAL

NEURAL NETWORK DESIGN

Research in the area of neural brain functions already

started centuries ago i.e. with the probabalistic model

of a perceptron (Rosenblatt, 1958). Rosenblatt cre-

ated the fundament on which Feed-Foreward neural

Networks (FFNN) are based. In general, they are

based on learned datasets of what is wanted to have

coming out of the network. This is also called super-

vised learning (Jordan and Rumelhart, 1992).

On the other hand, unsupervised learning is com-

monly understood as putting inputs into the network

and let the network ﬁll out the missing parts based

on what was previously trained (Hofmann, 2001).

Boltzmann machines (Hinton and Sejnowski, 1986),

Markov chains (Kani and Ardehali, 2011) and Hop-

ﬁeld networks (Hopﬁeld, 1982) are more similar to

the SLA model because each node is related to all

other nodes and also some basic responsibilities are

deﬁned such as input, output and restrictions to some

nodes but not all of them like in the SLA model. Any

of the mentioned conceptual models and also other

more modern Encoders (Gregor et al., 2015) are all

based on the inputs and the expected outcome. This

is the main difference to the SLA process. The SLA

model processes each constellation of the network

without an expected outcome. It transforms the net-

work into domain clusters to get outcomes related to

speciﬁc universals which are taxonomy-based signif-

icant to the provided inputs and of course take any

changing aspect into account by the transformation

step and relate it to all of the other aspects in a prob-

abalistic way during the association step of the SLA

processing chain.

5 APPLICATION TO REAL

WORLD EXAMPLES

The SLA process was tested in two different envi-

ronments. First, it was tested if it can produce rele-

vant outputs in form of recommendations in the ﬁeld

of technical mechanics and integrated circuit parts of

the public transport industry. The second environ-

ment, football game predictions, was selected to be

a completely different context compared to the ﬁrst

simulation. The inputs and training data as well as

the output of the second simulation was selected to

be completely different. Football game predictions

were considered as being easier to adapt and to test

and are easier to understand than technical concepts

in the mechanical and electronical industry.

5.1 Automatically Finding Obsolescence

Management Solutions

According to Bartels (Bartels, 2012), integrated cir-

cuits and electronical parts are obsolete if the industry

is not able to have them delivered anymore or they are

not needed anymore e.g. because of law restrictions

or technological evolution. Many railway companies

have the same problems and need to ﬁnd a way to re-

place obsolete parts to keep public transport operable

longer then estimated without obsolescence manage-

ment. At this point, obsolescence management fo-

cuses on the lifetime extension of any kind of techni-

cal construction.

Simulation design I.

For the communication between railway companies

on the subject of obsolescence management, an inter-

net platform has been developed. Currently about 50

people from 20 different railway companies commu-

nicate on different obsolescence management tasks

through the platform every day. Figure 3 shows how

the obsolescence cases have grown over time. Im-

portant to know is that the cases come from different

modes of transportation e.g. trains, metro, tram, bus.

Focusing all modes of transportation simultaneously

is useful because parts and internal components can

be identical across different modes of transportation,

e.g. train, bus, metro, etc. However, the aim of the

SLA process in this area is to predict which parts can

be replaced with a commonly accepted solution au-

tomatically at the time a new obsolescence case was

entered. Entered cases are saved according to the in-

ternational system of railway labeling EN 15380-2,

thus the matching algorithm can learn based on in-

ternational standard input sequences provided by user

inputs. The procedure of doing obsolescence manage-

ment is also conform with the international standard

DIN EN 62402. This enables the algorithms to ana-

lyze user behavior and learn which recommendations

are most relevant to the user at a speciﬁc time point.

Methods applied I.

To accomplish the main tasks in this area the learn-

ing algorithm learns from user inputs and produced

Real World Examples of Agent based Decision Support Systems for Deep Learning based on Complex Feed Forward Neural Networks

Figure 3: Obsolescence Management Data Growth.

data entered by the user or generated through the SLA

process for a speciﬁc environment or product fam-

ily. If the user enters a complete obsolescence case

with its solution included, the SLA process searches

for similarities between different other obsolescence

cases and matched versions, part numbers and other

criteria applicable for the identiﬁcation, location and

the speciﬁc environment of a given part at collection

time. The association step binds all collected items

together into environments or product families in or-

der to operate on them in the next step. At operation

time each of the associations were mapped to different

speciﬁc patterns previously stored and related to the

speciﬁed product families and component interfaces.

If nothing was found, the collections and associations

are transformed using positive and negative feedback

loop algorithms combined with possibly missing data

to force a pattern match. Otherwise if something sig-

niﬁcant were found, the grade of relevance for the so-

lution gets stored for further usage in next iterations.

Results I.

The simulation has shown that in the ﬁeld of obso-

lescence management the SLA process was less use-

ful in the speciﬁed area of component part mappings.

We ﬁgured out that we can make good suggestions to

the user which components could ﬁt together, unfor-

tunately there was not enough data to work on rele-

vant pattern matching. We have not been able to con-

struct as much relevant data across the different rail-

way environment and product families to get signiﬁ-

cant mapping patterns which are pointing on solutions

the industry can use in practice. The most difﬁcult

area of transforming and operating data was the in-

compatible design of integrated components. At con-

struction time constructors did not consider the com-

patibility across the interfaces of other, mostly com-

petitive integrated circuits or mechanical components

of other railway or public transportation vehicle part

creators. This kind of investigation seems to have

much better results if creators are forced to design

compatible interfaces by governance, law or interna-

tional standards.

5.2 Football Game Prediction

Another approach for SLA process application is the

area of football game statistics to predict a football

teams performance compared to other football teams

during an ofﬁcial football match across nearly 60 dif-

ferent leagues around the world. This simulation was

selected out of many different options because of the

grade of difference to the study previously described,

to test the applicability of the SLA process for differ-

ent ﬁelds of data computation and different kind of

predictions or expected outcomes.

Simulation Design II.

The statistical baseline for game predictions of the

last three football seasons in the learning part of the

SLA process was retrieved from a public available

database via http://www.sofascore.com (Sofa). The

Application Programming Interface (API) of Sofa

COMPLEXIS 2017 - 2nd International Conference on Complexity, Future Information Systems and Risk

Figure 4: Bet Filtering.

provides all relevant data e.g. on matches, scores,

teams and players of several football leagues. The

SLA Model was designed to work in the same way as

a professional football expert would bet. Building up

domains for leagues, teams, and regional history, then

collecting data about past game results of the com-

peting teams (Step 1 Collection), players, judges and

observing the motivation regarding the current league

rank of relevant teams. Operations (Step 3 Operation)

have been applied to calculate a player score for each

possible team formation during the match (Step 2 As-

sociation). Each constellation was saved (Step 5 Stor-

age) as a relevant team score (Step 4 Deﬁnition) and

was initiated as an own transformable network (Step 6

Transformation). Any relevant information was then

calculated for each of the possible team formations

to calculate a player score for each player in differ-

ent ﬁeld positions and their relevant opponents in the

attack-, middle-, and defense-ﬁeld. The results have

been saved in a meta network in which all predictions

for one football match have been placed. Any constel-

lation of the variables did predict a speciﬁc chance of

competing the opponent in a simple value based on

randomly generated data for the speciﬁc match situa-

tions the players usually have.

Methods applied II.

To predict which team will win in a speciﬁc football

match, different methods have been applied. The pre-

diction which team will win was interpreted out of

a percentage rate of consensus for each of the meta

network values and generated data on which they are

based on. The illustration in Figure 4 shows accepted

bets and bets which have been ﬁltered out because

they did not match the createria to have more than

80% meta network consensus in a rising amount of

different team constellations and in-match player sit-

uations. Predictions cannot be placed if the relevant

data is not available. Usually the player formation

data is available 30 minutes before the game starts.

An automated algorithm extracts the data from Sofa

Figure 5: Bet Quote Results.

and hands it over it to the SLA System. Within sec-

onds it is calculated if the bet is ﬁltered out or will

enter the betting queue. There are three options for

placing a bet as illustrated in Figure 5. The SLA Sys-

tem was learned to ﬁnd out the best betting option on

a previously calculated range based on the Gaussian

bell curve concerning the bet quotes, player and team

constitution as well as different approaches to reduce

wrong predictions because of outlying values. A bet

was always only placed if all meta networks gave a

consensus of at least 80%, otherwise the bet for the

match was rejected. Initial simulations indicated un-

reliable predictions, extrem outliers or many missing

data which is why for several leagues it was decided

not to bet at all.

Results II.

Figure 7 shows how the simulation did predict a con-

crete income for a comparable low amount of possi-

ble bets at a computational low risk. The discrepancy

between theory and practice regarding football game

predictions, based on the SLA algorithm learned on

real seasonal data from the past, becomes clear when

comparing to the bet reality illustrated in Figure 6,

where earnings decrease with increasing number of

bets. Thus, football game predictions are harder to

predict if they come from complex automated algo-

rithms. The analysis of the outcomes has shown that

there were huge amounts of football matches where,

according to the statistics, stronger teams lost the

match. Compared to the Obsolescence Management

study, the main differences are the reliability and the

amount of patterns found. In football game predic-

tions there where signiﬁcantly less patterns found and

these where also less reliable compared to the area of

recommendation engines in the ﬁeld of public trans-

portation technology.

6 DISCUSSION

Analytical Data Processing for recommender sys-

tems (Manouselis et al., 2011) like those systems de-

veloped for e-commerce decision making (Kim and

Srivastava, 2007) or machine translation (Bahdanau

et al., 2014) or acustical noice reduction (Srivastava

Real World Examples of Agent based Decision Support Systems for Deep Learning based on Complex Feed Forward Neural Networks

Number of bets

Earnings

Figure 6: Bet Simulation.

Number of bets

Figure 7: Bet Reality.

et al., 2014) already use machine learning algorithms

based on neural networks. They are all extremely spe-

cialized and cannot be used in other environments as

they have been made for. Ubiquitous computing or

at least the Social Internet of Things (SIoT) (Atzori

et al., 2012) will hold many autonomous complex

systems with its own business logics and their own

intelligence in a peer-to-peer environment of similar

autonomous systems sharing information all together

in different environments. Environments change very

fast or need to cover unpredictable changes. Any ma-

chine learning process should be able to learn from

the inputs and outcomes and thus to reconﬁgure them-

selves to be able to react on unexpected environmen-

tal changes. In comparison to other research, the SLA

process, based on biological neuron functions is sim-

ilar to the approaches of combining artiﬁcial intelli-

gence with widely approved biological and chemical

processes like DNA Sequencing (Ezziane, 2006) or

the classiﬁcation of protein structures (Suhrer et al.,

2009) as both approaches extract knowledge from an

enormous amount of data. Further investigations are

necessary to also cover following aspects:

First, the ”State of Mind” concept, how different

Neuronal Networks can be used to represent some-

thing similar to the ”state of the mind” for a speciﬁc

recognition or the polarity between sentiments (Po-

ria et al., 2014). Second, ”Priming” concept, how

the element association can be marked as complete,

abstract, stable or has a signiﬁcant convergence to a

speciﬁc state deﬁned previously of being important

for the current reasoning process based on the con-

sensus of positive and negative feedback loops.

Third, ”Balancing” concept between positive and

negative feedback (nF <> pF) to store a speciﬁc deﬁ-

nition (D) for later use. Fourth, theoretical complexity

and performance through efﬁcient cache, transience

and garbage collection models are not focused yet.

Last, ”Avoiding” concept in order to shrink, expand or

test the neural networks in order to rise effectiveness

of the learning function (e.g. reducing errors faster)

(Srivastava et al., 2014).

7 CONCLUSION

The combination of both, the SLA and the Dual

Thinking process was considered to be a good ﬁt

for making progress in simulating human thinking

processes for technology enhanced learning in the-

ory. But both of the studies exempliﬁed in this paper

show that something of higher signiﬁcance is missing

to compare computational simulation processes, like

SLA with natural learning. Machine learning algo-

rithms, especially in the manner of imagination and

reasoning, will improve many modern life scenarios

in the near future. Decision making based on the be-

havioral history of complex environments like in the

SLA process could take a signiﬁcant part of it. These

decision making processes deﬁnitely need further in-

vestigation and deeper statistical analysis for positive

and negative feedback relevances. The level of sig-

niﬁcance need to be saved in the knowledge map as

well. Only in this way it is possible to know the share

of a knowledge map fractal on the decision making

process. This enables change prediction in the trans-

formation processing step and could lead to more ef-

ﬁcient conclusions. It seems there exists a necessity

for statistical baselines and standardized taxonomy in

each knowledge transformation step. Operations, se-

lections, associations and transformations should be

based on statistical data analysis and need more data

as used in both approaches exempliﬁed in this re-

search.

REFERENCES

Atzori, L., Iera, A., Morabito, G., and Nitti, M. (2012).

The social internet of things (siot)–when social net-

works meet the internet of things: Concept, archi-

tecture and network characterization. Computer Net-

works, 56(16):3594–3608.

Bahdanau, D., Cho, K., and Bengio, Y. (2014). Neural ma-

chine translation by jointly learning to align and trans-

late. arXiv preprint arXiv:1409.0473.

COMPLEXIS 2017 - 2nd International Conference on Complexity, Future Information Systems and Risk

100

Bartels, B. (2012). Strategies to the prediction, mitigation

and management of product obsolescence. Wiley.

Cooper, G. F. (1990). The computational complexity

of probabilistic inference using bayesian belief net-

works. Artiﬁcial intelligence, 42(2):393–405.

Ezziane, Z. (2006). Applications of artiﬁcial intelligence in

bioinformatics: A review. Expert Systems with Appli-

cations, 30(1):2–10.

Grady, L. J. and Polimeni, J. (2010). Discrete calculus:

Applied analysis on graphs for computational science.

Springer Science & Business Media.

Graves, A. (2013). Generating sequences with recurrent

neural networks. arXiv preprint arXiv:1308.0850.

Graves, A. and Jaitly, N. (2014). Towards end-to-end

speech recognition with recurrent neural networks. In

Proceedings of the 31st International Conference on

Machine Learning (ICML-14), pages 1764–1772.

Graves, A., Liwicki, M., Fern

andez, S., Bertolami, R.,

Bunke, H., and Schmidhuber, J. (2009). A novel

connectionist system for unconstrained handwriting

recognition. Pattern Analysis and Machine Intelli-

gence, IEEE Transactions on, 31(5):855–868.

Graves, A., Mohamed, A.-r., and Hinton, G. (2013). Speech

recognition with deep recurrent neural networks. In

Acoustics, Speech and Signal Processing (ICASSP),

2013 IEEE International Conference on, pages 6645–

6649. IEEE.

Gregor, K., Danihelka, I., Graves, A., Rezende, D. J.,

and Wierstra, D. (2015). Draw: A recurrent neu-

ral network for image generation. arXiv preprint

arXiv:1502.04623.

Hinton, G. E. and Sejnowski, T. J. (1986). Learning and re-

leaming in boltzmann machines. Parallel distributed

processing: Explorations in the microstructure of cog-

nition, 1:282–317.

Hofmann, T. (2001). Unsupervised learning by probabilis-

tic latent semantic analysis. Machine learning, 42(1-

2):177–196.

Hopﬁeld, J. J. (1982). Neural networks and physical sys-

tems with emergent collective computational abili-

ties. Proceedings of the national academy of sciences,

79(8):2554–2558.

Ji, L., Liu, Q., and Liao, X. (2014). On reaching group

consensus for linearly coupled multi-agent networks.

Information Sciences, 287:1–12.

Jordan, M. I. and Rumelhart, D. E. (1992). Forward models:

Supervised learning with a distal teacher. Cognitive

science, 16(3):307–354.

Kani, S. P. and Ardehali, M. (2011). Very short-term wind

speed prediction: a new artiﬁcial neural network–

markov chain model. Energy Conversion and Man-

agement, 52(1):738–745.

Kim, Y. and Srivastava, J. (2007). Impact of social inﬂuence

in e-commerce decision making. In Proceedings of

the ninth international conference on Electronic com-

merce, pages 293–302. ACM.

Kisch, H. R. and Motta, C. L. R. (2015). Model of a neu-

ron network in human brains for learning assistance

in e-learning environments. In Proceedings of the 7th

International Conference on Computer Supported Ed-

ucation, pages 407–415.

Krizhevsky, A., Sutskever, I., and Hinton, G. E. (2012). Im-

agenet classiﬁcation with deep convolutional neural

networks. In Advances in neural information process-

ing systems, pages 1097–1105.

Le, Q. V., Jaitly, N., and Hinton, G. E. (2015). A simple

way to initialize recurrent networks of rectiﬁed linear

units. arXiv preprint arXiv:1504.00941.

Liao, X. and Ji, L. (2014). On pinning group consensus

for dynamical multi-agent networks with general con-

nected topology. Neurocomputing, 135:262–267.

Luong, M.-T., Sutskever, I., Le, Q. V., Vinyals, O., and

Zaremba, W. (2014). Addressing the rare word prob-

lem in neural machine translation. arXiv preprint

arXiv:1410.8206.

Manouselis, N., Drachsler, H., Vuorikari, R., Hummel, H.,

and Koper, R. (2011). Recommender systems in tech-

nology enhanced learning. In Recommender systems

handbook, pages 387–415. Springer.

Ochs, P., Ranftl, R., Brox, T., and Pock, T. (2016). Tech-

niques for gradient-based bilevel optimization with

non-smooth lower level problems. Journal of Math-

ematical Imaging and Vision, pages 1–20.

Poria, S., Cambria, E., Winterstein, G., and Huang, G.-B.

(2014). Sentic patterns: Dependency-based rules for

concept-level sentiment analysis. Knowledge-Based

Systems, 69:45–63.

Rosenblatt, F. (1958). The perceptron: a probabilistic model

for information storage and organization in the brain.

Psychological review, 65(6):386.

Scheffer, M., Bascompte, J., Bjordam, T. K., Carpenter,

S. R., Clarke, L., Folke, C., Marquet, P., Mazzeo, N.,

Meerhoff, M., Sala, O., et al. (2015). Dual thinking

for scientists. Ecology and Society, 20(2).

Sporns, O. and Koetter, R. (2004). Motifs in brain networks.

PLoS Biology, 2(2):e411.

Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I.,

and Salakhutdinov, R. (2014). Dropout: A simple way

to prevent neural networks from overﬁtting. The Jour-

nal of Machine Learning Research, 15(1):1929–1958.

Suhrer, S. J., Wiederstein, M., Gruber, M., and Sippl,

M. J. (2009). Cops a novel workbench for explo-

rations in fold space. Nucleic acids research, 37(suppl

2):W539–W544.

Sutskever, I., Vinyals, O., and Le, Q. V. (2014). Se-

quence to sequence learning with neural networks. In

Advances in neural information processing systems,

pages 3104–3112.

Real World Examples of Agent based Decision Support Systems for Deep Learning based on Complex Feed Forward Neural Networks

101