You are on page 1of 31

Running head: BIOLOGICALLY INSPIRED REACTIVE PLANNER

Instinct: A Biologically Inspired Reactive Planner for Embedded Environments

[Name]
[Professor]
[Date]
BIOLOGICALLY INSPIRED REACTIVE PLANNER 1

Table of Contents
Literature Review............................................................................................................................ 3

1.1 Instinct ................................................................................................................................... 3

1.2 Artificial Neural Network – Biological Inspiration .............................................................. 4

1.2.1 General Structure of Neuron........................................................................................... 5

1.2.2 Machine Learning ........................................................................................................... 6

1.3 Robotics Operating System ................................................................................................... 7

1.3.1 Dynamic Programming................................................................................................... 8

1.4 Computer Simulation ............................................................................................................ 8

1.4.1 Hybrid Controlled Architect ........................................................................................... 9

1.5 Mobile Robotics .................................................................................................................. 10

1.6 Cognitive Psychology within Robotics ............................................................................... 11

1.7 Reactive Planners ................................................................................................................ 12

1.7.1 Dynamic Hierarchical Compact ................................................................................... 13

1.8 Representation of knowledge .............................................................................................. 14

1.9 Cognitive Architectures ...................................................................................................... 15

1.9.1 Agent ............................................................................................................................ 15

1.9.2 Autonomous Agent ....................................................................................................... 15

1.9.3 Cognitive architecture................................................................................................... 15

1.9.4 Behavior........................................................................................................................ 16

1.9.5 Intelligence ................................................................................................................... 16

1.9.6 Knowledge .................................................................................................................... 16

1.10 Characteristics of a cognitive architecture ........................................................................ 16

1.11 Controlled Architectures ................................................................................................... 18

1.12 Behavioral Design ............................................................................................................. 19


BIOLOGICALLY INSPIRED REACTIVE PLANNER 2

1.13 Natural Selection in Ecology............................................................................................. 21

1.14 The second theorem of natural selection ........................................................................... 21

1.15 Exclusive selection of behaviors – Utility function method ............................................. 22

1.16 Subsumption Architecture ................................................................................................. 23

1.17 Expert Systems .................................................................................................................. 25

References ..................................................................................................................................... 26

List of Figures
Figure 1: Artificial Neural Network ............................................................................................... 5
Figure 2: Machine Learning Techniques ........................................................................................ 6
Figure 3: Graphs based on β -similarity, β = 0.3. ......................................................................... 14
Figure 4: Framework of Cognitive Architecture .......................................................................... 18
Figure 5: Reactive Model.............................................................................................................. 20
Figure 6: Subsumption Architecture ............................................................................................. 23
BIOLOGICALLY INSPIRED REACTIVE PLANNER 3

Literature Review

1.1 Instinct
Wortham, Gaudl, and Bryson [1] have described instinct planner as a novel reactive planner, which
was biologically inspired based on those behaviors that follow the methodology of robotics, as
well as its components of the reactive planner. This component relates to the implementation of
the POSH planner. This planner comprises several important enhancements that support the design
of the plan, as well as debugging of the runtime. It was specifically created for low power
processors and contained a tiny footprint of memory.

The language in which this planner was written is C++. In addition to this, it runs effectively on
both the environments of Microsoft VC++, as well as ARDUINO (ATMEL AVR). This instinct
planner was implemented to study the transparency of AI within a low-cost maker robot. Wortham
et al. [1] further stated that plans could be authored by utilizing different tools, which include a
promising design of the visual language that is currently applied by utilizing the drawing package
of DIA.

Instinct planner can be described as a reactive planner that was created based on POSH, which was
proposed by Bryson. It comprise of numerous enhancements that were gathered from recent
research papers that were extended on POSH. Moreover, it also include several ideas from other
approaches of planning, one of which is Behavior Trees. The instinct planner also comprised of
full application of drive execution optimization. DEO prevents the entire plan tree to be searched
at every cycle of the plan, because this process is highly expensive. It also make sure that focus is
maintained on the tasks at hand. This process resembles loosely to the functioning that is observed
in consciousness attention, which is mostly witnessed in the nature [1].

To create an appropriate planner that efficiently operates under those environment where resources
of the working memory is extremely limited, substantial efforts on the design has been
implemented within the architect of the instinct planner’s memory management. Furthermore,
there are six different buffers of the memory under this design that are holding fixed elements
related to the length of the record for every type of element within the plan. These six buffers are
Actions, Action Pattern Elements, Action Patterns, Competence Elements, Competencies and
BIOLOGICALLY INSPIRED REACTIVE PLANNER 4

Drives. A single collection of drive can be found in one instance associated with the instinct (i.e.,
the root or source of the plan) [2].

For the values of Boolean, under each element of the plan, every byte is divided into bit fields.
Moreover, to prevent variation within the length records, normalization of the data is made across
all the elements. In other words, this means that the parent competence’s ID is held with the
competence elements. However, the competence by itself does not hold any ID of each of the
competence elements’ child. In addition to this, it is necessary to conduct a search during run time
in order to determine which competence elements has a link with the provided competence. Hence,
the planner makes some sacrifices under the search time to attain a more considerable
representation of the compact memory. Fortunately, the search does not take a lot of time, because
the competence elements are kept under a single buffer of memory, having length records that are
fixed. The testing highlights that the time which is required by this searching was insignificant as
compared to the robot’s plan cycle rate.

1.2 Artificial Neural Network – Biological Inspiration


According to Wortham and Bryson [2] science has been inspired by nature in a variety of
environments. It can be seen, for example, how aircraft use aerodynamic principles taken from the
way birds stay in the air, the frog fins used by divers to improve the movement in the water are
inspired by that animal, the functioning of the Cameras have been perfected through the imitation
of the eyes, among other examples.

Nature provides a powerful tool to learn from their mistakes through natural selection and comes
to designs that generate great efficiency and adaptation to their environment and that feature makes
many of them become great ideas technologically speaking [3].
BIOLOGICALLY INSPIRED REACTIVE PLANNER 5

Figure 1: Artificial Neural Network


In the field of artificial intelligence there are several examples of how natural models have
been adapted to create algorithms that allow us to learn from their mistakes and become more
efficient in their task. These algorithms are particularly useful for solving tasks in which it is not
possible to find a mathematically pure solution. An example of the above is artificial neural
networks [4].

1.2.1 General Structure of Neuron


Low, Leow and Ang [5] stated that the neurons are connected to each other by means of
connections called synapses which are electrical exchanges between the synaptic button of one
neuron (transmitter or presynaptic) and the dendritic button of another (receptor or post synaptic).
The electric discharge travels from the body or soma of a neuron to the Axon which acts as an
electric cannon through the neurotransmitters.

Inspired by this biological mechanism, McCulloch & Pitts created an artificial model of a
neuron in 1943. Later this idea was taken up by Rosenblatt who created a structure called
perceptron. The perceptron is a structure which has one or several inputs, an activation function
and an output [6].
BIOLOGICALLY INSPIRED REACTIVE PLANNER 6

1.2.2 Machine Learning


Murphy [7] stated that in order for the neural network to "learn", Machine Learning
techniques (see below figure) are used, which can be of a varied nature, however the most common
are those of supervised learning, in which data previously tagged are given as input to the
algorithm. A human "supervises" the training by telling the algorithm how the information should
be interpreted. For example, when people are trying to build a neural network capable of
classifying whether a fruit is an orange or a banana from its description.

Figure 2: Machine Learning Techniques

Each of the entries represents a characteristic of the object. People can design the network
so that the entrance one activates if the fruit has yellow color and the entrance two is activated
when the fruit is round. The intention of the people is that when a banana is described, entry one
is activated and entry two does not, and when an orange is described, the reverse process occurs.
So for the network to learn to differentiate an orange and a banana, people will give examples to
the algorithm in which when the first input is activated and not the second the correct answer is
banana and when the reverse process occurs the answer is orange [8].

According to Prescott, Bryson and Seth [9], this process is automated by means of error
propagation algorithms comparing the output value thrown by the neural network with the output
value of the tagged data, calculating the difference and considering it as an error. Then the weight
BIOLOGICALLY INSPIRED REACTIVE PLANNER 7

of that error is distributed so that the neurons "adjust" their connections until they have the desired
results. So the learning process in neural networks, is essentially about finding the appropriate
weights for each of the neuron links.

Mathematically speaking what happens as the information passes from one layer to the
other, is a composition of functions, which can be extremely complex, so describe what is
happening in each of the layers is an interesting challenge, without embargo. Behind this simplified
neuron structure, a series of increasingly complex mechanisms have been proposed that have
managed to perform surprising tasks, even performing better than humans in different tasks [10].

Brooks [11] argues that the world of neural networks has become increasingly complex
and extensive, but the direct inspiration of these algorithms comes from of the biological neuronal
model. Nowadays there is an interesting exchange between the neurosciences and the computer
sciences which are learning from each other and have brought a very exciting new era in machine
learning algorithms, in which the machines can learn thanks to the imitation of our brain and maybe
the way machines learn can help us to unravel things that humans still do not know about their
biology.

1.3 Robotics Operating System


In traditional sense, ROS cannot be considered as an operating system within the
scheduling and management process. Rather if offers a structured layer of communications above
the operating systems of the host within the compute clusters that are observed to be
heterogeneous. Writing or creating a software for robots is a challenging task, especially under the
conditions where the scope and scale of robotics is witnessing a consistent growth. It is observed
that various types of robots have substantially differing hardware, which makes the reuse of the
code nontrivial. Moreover, the substantial size of the needed code can be intimidating, as it is
required to hold a deep stack which starts from the driver level software and moves up through the
abstract reasoning, perception and so forth.

According to Gaudl and Bryson [12], the history of robotics goes back almost to the first
human civilizations. The Egyptians and the Greeks began to design the first automatisms that
vaguely resembled the robots nowadays. In most cases these mechanisms were inspired by animals
and human beings (although they were rarely anything other than mere entertainments). This trend
BIOLOGICALLY INSPIRED REACTIVE PLANNER 8

continued throughout centuries until the industrial revolution it was thought that "robots" could be
used to increase production of workers (or even replace them).

From 1946, with the advent of computer science, robotics began its first real development
era, where the robots became more and more useful and precise, becoming part of the majority of
production processes industry standards. In a next step the robots began to emerge from the
industrial environment to be used in a large number of fields: Exploration, recreational
applications, military purposes, etc. At this point the engineers realized that much of the
technology they had inherited from the application industrial conditions of the robots did not work
for these other fields, which presented problems that until now had not been found and that seemed
unsolvable with the techniques that had been used until now [13].

1.3.1 Dynamic Programming


Dynamic programming has been considered as one of the powerful approach, which can
be utilized to tackle various problems in time O (n 3), as well as O (n 2) for which a naïve
methodology would consume exponential time. Dynamic programming can also be considered as
a more general approach to solve issues, much similar to ‘divide and conquer’ but the sub problems
under this approach are usually found to overlap each other. In both algorithms’ design, as well as
control theory, the optimality principles has reflected a robust constraint within the set of candidate
solutions to optimize the problems. Moreover, the methods of dynamic programming are restricted
to state spaces that are related to only few domains. For instance, the tree of rapidly exploring
random was used to explore algorithms to increase the searching speed of high-dimensional
spaces, which have both differential constraints and global constraints.

1.4 Computer Simulation


Computer simulation allows recreating the behavior of organisms in the environments or
niches in which they live, modeling their sensors, the processing of information they carry out and
the tasks they perform. Once the simulation is done, it can be implemented in physical or virtual
robots or computational agents that have access to information sources and are able to verify and
manipulate the information obtained from these sources according to the queries raised by the
users. in order to obtain the expected resources. The sources of information can be of many types,
including, for example, databases and information on the Internet [14].
BIOLOGICALLY INSPIRED REACTIVE PLANNER 9

According to Brom et al. [15] the study of animal behavior and neurobiological
experiments can serve as a guide for the design and construction of artificial systems that achieve
autonomous and adaptive behavior: When an animal confronts its environment, its behavior is
continuously adjusted to face situations, needs and conditions of the environment both internally
and externally. For example, feeding, fighting, fleeing and mating are four basic impulses (known
in biology as the 4 Fs) that the body can face and determine their behavior, in this case, survival.

In the processing of information carried out by organisms and artificial systems, the brain
(the controller), the body (internal and external interactions) and biunivocal relationships with their
environment are involved. In artificial agents this tripartite relationship is represented by the
control architecture that includes: the perception and selective action on the environment; the
processing of internal and external information and; the sensors, actuators or effectors [16].

The selection of the most appropriate behavior by the agent is a subject known as "action
selection" and is very important when the agent is expected to exhibit intelligent and adaptive
behavior that considers the prediction and evaluation of the consequences of their actions. The
selection mechanism is placed in a layer of the architecture or is distributed in smaller, more
manageable and flexible systems that allow the arbitration and prioritization of behaviors and that
manage the maintenance, inhibition and alternating behaviors when they have multiple conflicting
objectives.

1.4.1 Hybrid Controlled Architect


According to Breazeal and Scassellati [17], hybrid control architecture combines the
advantages of reactive and deliberative approaches, while decreasing its individual drawbacks, but
its computational complexity is greater. The deliberative agent generates in a static way plans that
will be executed in real time, reacting to the environment. Finally, the structure of a behavioral
control architecture is implemented from a collection of independent behaviors that are self-
organized to carry out the tasks entrusted to them.

Nilsson [18] stated that of these, reactive architecture is very popular for its simplicity,
because it responds immediately to the stimuli of the environment, even when they do not have
the capacity for anticipation, memory or learning. Examples of reactive architectures include Gray
Walter turtles in the 1950s, Valentino Braitenberg vehicles and Rodney Brooks creatures of the
1980s. The turtles of Walter are speculative machines that explored their environment, by means
BIOLOGICALLY INSPIRED REACTIVE PLANNER 10

of light and contact sensors, as well as two engines one to propel it and the other to be able to guide
it in the environment. The processing was done with an analog bulb circuit. Turtles exhibit
complex behaviors (seeking a light, moving away from it and avoiding obstacles).

Braitenberg vehicles use simple excitatory and inhibitory connections between sensors and
motors, producing behaviors such as aggression, fear, and love exploration. The Brooks creatures
are based on the subsumption architecture, constituted by layers that work in parallel and that are
organized in a prioritized way by means of a competitive coordination mechanism. The behaviors
show a high degree of intelligence without, in fact, a central processor.

Currently, according to Wortham, Theodorou, and Bryson [19] behavioral architecture is


the basis of the design of many intelligent agents, with which the agent chooses and coordinates
the most appropriate behavior according to the stimuli received and the context in which it is
located. For example, Somerville and Casey [20] present a behavioral control system for the
interaction of the human being with customizable social robots that adapt their behaviors to the
profiles of the users and to the responses during the human-robot interaction. The programming of
an agent's behaviors can be inspired by the way in which animals resolve the conflict of selecting
a behavior at the moment when several functional units enter into competition for a limited set of
actions and only one functional unit takes the control of motor exits.

1.5 Mobile Robotics


Scientific interest in mobile robotics has been growing rapidly as evidenced by the large
number of publications and scientific meetings held in recent years. One of the main reasons for
this growth, apart from the enormous potential of industrial applications already mentioned, is that
autonomous mobile robots are an important means of research and development of theories about
the behavior of living beings by providing a possibility of experimentation.

Seth [21] stated that for intelligent mobile agents, whether robots or living beings, the
ability to locate, orient and plan paths is of the greatest importance. This ability to navigate requires
not only a reactive behavior, but also the ability to interpret information about global objectives,
intelligent interaction with the environment, and the ability to recover from errors. Navigation is
therefore an ideal medium for checking and researching theories about intelligent behavior. That
is, the mobile robots serve the sciences dedicated to the study of the mind as a test bank of their
theories.
BIOLOGICALLY INSPIRED REACTIVE PLANNER 11

This influence is reciprocal if we consider that the study of the behavior of living organisms
can help us in the study of the behavior of artificial mechanisms. The principles extracted from the
large amount of data stored on animal behavior can be used to improve the design of the robots.
For the designer of autonomous mobile robots, the ontological point of view is interesting,
according to which the robot must occupy its ecological niche within a world shared with other
agents and from which it must reach its mission successfully.

As has happened so many other times in the history of science, nature is also in this field a
powerful source of inspiration. The robots of this new generation must be able to share more
closely the activities of human beings and for that they must be able to function in an evolving
world to which they must adapt, perhaps using the same means as human beings, through a labor
continuous learning: by experimentation, by imitation, by the use of didactic materials, or with the
help of an external agent acting as a teacher.

1.6 Cognitive Psychology within Robotics


The project of design and construction of a mobile autonomous robot implies the
coordinated use of specific knowledge coming from several fields of Technology: Mechanics,
Electricity, Electronics, and Computing. The search of research works carried out in the field of
practical realization of systems capable of developing control actions with the adaptability and the
required problem solving capacity leads us directly to the studies that are currently carried out on
Cognitive Architectures. This field of study belongs to the research program defined by Cognitive
Science.

Varela, Thompson and Rosch [22] stated that cognitive science tries to elucidate the
behavior of the mind, sometimes explicitly, and sometimes implicitly, treating it as computation,
not necessarily of the type carried out by digital computers, but of a broader type, as is the theory
computer mathematics. Similarly, according to Clark [23] cognitive science is a domain defined
by all objects capable of knowing, minds and computers. The general science of systems capable
of performing computations based on representations." It can be said that the current research on
mobile robotics is related to the development of several technical and scientific disciplines:
Mechanical Engineering, Electricity, Electronics, and Information and also Cognitive Science that
is, Psychology, Artificial Intelligence, Philosophy, Linguistics, Anthropology, Neurophysiology
and Education Sciences.
BIOLOGICALLY INSPIRED REACTIVE PLANNER 12

Volos, Kyprianidis, and Stouboulos [24] argued that the approach to the contents of the
area of the design and development of autonomous mobile robots is possible from several
disciplines. From the point of view of Industrial Engineering, where there is a growing demand
for more flexible robots, with greater freedom of movement, more adaptable to changes in the
environment, with greater capacity for learning.

From the world of Artificial Intelligence where the mobile robot appears as an active agent
that, unlike the more traditional applications of this discipline focused on the development of
symbol processors, must be developed in a real environment captured through sensors and that
must move in a controlled manner. And from the point of view of Cognitive Psychology where
mobile robots are, perhaps, "objects with minds" possessing characteristics that make them an
excellent laboratory for the study of different proposals of cognitive architectures.

It is clear that in nature there are a large number of examples of beings that, located at a
medium level of the evolutionary scale, develop a behavior that meets the requirements listed in
the previous paragraph. These requirements taken to its most demanding version are more than
solved by the intelligence capabilities of human beings. Faced with such a broad spectrum of study
possibilities, the designer of an autonomous mobile robot is currently faced with a first barrier that
is difficult to overcome when, in the Analysis Phase, they try to compile the available knowledge
that may be useful.

Williams [25] stated that for the realization of the construction project of a robot with
specific requirements. In fact, one of the main problems faced by the research and development of
autonomous mobile robots is the lack of a common scheme in which it is possible to integrate the
results achieved by different research teams, which can serve as a reference to compare work, and
that allows to reuse the efforts made from the different scientific disciplines that participate in this
field of study.

1.7 Reactive Planners


It has been stated in one study [26] that in many practical problems it is necessary to create
hierarchies of objects such as the construction of biological taxonomies and the detection in a
continuous flow of news of the topics and the events in which they are decomposed. Note that in
some of these examples, it is also necessary that the hierarchical algorithm be dynamic due to the
nature of the collections to be grouped. Many hierarchical grouping algorithms have been
BIOLOGICALLY INSPIRED REACTIVE PLANNER 13

developed, among them we can mention the Average-link, Complete-link, Single-link, ICT,
Bisecting K-Means, as well as UMASS. These algorithms require the complete collection of
objects to perform the grouping and it is inefficient to use them in problems where the collections
are dynamic. In addition, the few incremental hierarchical algorithms that exist, such as GALOIS,
the Charikar algorithm and DC-tree, present several of the following drawbacks:

1. They have an exponential temporal complexity with respect to the dimension of the objects,

2. They need to determine a priori the number of groups to obtain,

3. They require many parameters, or

4. The groups obtained have a high dependence on the order of presentation of the objects.

1.7.1 Dynamic Hierarchical Compact


Dynamic Hierarchical Compact algorithm is able to maintain less computational time unlike other
standard approaches or methods within the dynamic sets of data, while accomplishing a
comparable and even more effective quality of clustering [27]. This algorithm is able to maintain
a hierarchy of groups up to date and does not present any of the above limitations. Its computational
complexity is quadratic. To enable its use in large collections of objects, in this work we propose
its parallel version. The Parallel Dynamic Hierarchical Compact algorithm takes advantage of the
computing power available in a distributed memory computer. In the work we show the results
obtained in the grouping of four collections of documents in a cluster of PCs. This algorithm
manages to reduce the computation times to a certain number of processors.
BIOLOGICALLY INSPIRED REACTIVE PLANNER 14

Two objects whose similarity is greater than or equal to a certain threshold β (defined by
the user) are called β -like. If an object is not β -like with any other object it is called β -isolated.
It is called a graph of β -similarities to the graph where the vertices are the objects to be grouped
and there are edges between the objects β - similar, labeled with the value of their similarity. We
will call graph of maximum β -likeness (max-S), oriented graph where the vertices are the objects
to be grouped and there exists an arc from the vertex oi to the vertex oj if it is fulfilled that oj is
the object most β -similar to oi. Finally, denotation as U-max-S has been made in the graph
obtained from the max-S graph ignoring the orientation of its arcs (see figure below).

Figure 3: Graphs based on β -similarity, β = 0.3.

Given a hierarchy of groups, previously built by the algorithm, each time an object is added
or deleted it is necessary to update existing groups at all levels of the hierarchy. Before the arrival
of a new object, a unitary group is created and added to the lower level of the hierarchy. Then, the
graphs of β -similarity and max-S are updated and, in addition, the compact groups that constitute
the groups of the next level of the hierarchy are updated [28].

1.8 Representation of knowledge


According to Sowa [29] the expression representation of knowledge are establishing a
difference between the information contained in that knowledge and the form used for its
representation. In this work the hypothesis that the mind works with symbolic representations of
the world in which people live. Therefore to process that information and also to communicate
BIOLOGICALLY INSPIRED REACTIVE PLANNER 15

with each other people use representations. These representations may be different in the different
contexts in which the information is used. The representation adapts to the needs of the context.

When studying the best way to organize the cognitive architecture of an intelligent agent
who must face the complexity of their environment, people find themselves, as intelligent agents,
faced with the need to make use of adequate strategies to deal with the complexity of the problem
we are facing. Raising this uncomfortable component of recursion is characteristic of studies about
the functioning of the mind.

1.9 Cognitive Architectures


Since Newell and Simón published their work "Human problem solving" where they
describe a "General Problem Solver", a program that seemed to have general capabilities for the
proof of theorems, a certain number of architectures have been proposed that different approaches
try to solve all or some of the functions present in what we call an intelligent agent. Researchers
who have faced the development of a cognitive architecture have done so from different
motivations and points of view [30].

In studying the history of the development of cognitive architectures, focusing on the


application to the construction of autonomous mobile robots, we realize that in reality most of the
results obtained, far from having become obsolete, have provided solutions that have been
accumulated over time creating a scheme that although it has been increasing in complexity, has
been approaching without doubt to the resolution of the problem.

1.9.1 Agent
It is a system that can be differentiated from its environment with which it maintains a
direct interaction, sensing through sensors and acting through actuators.

1.9.2 Autonomous Agent


The activity of the agent does not require the intervention of an operator and their behavior
depends on their own experience.

1.9.3 Cognitive architecture


A cognitive architecture is an integrated structure of resources capable of supporting the
functions of an intelligent agent.
BIOLOGICALLY INSPIRED REACTIVE PLANNER 16

1.9.4 Behavior
An agent maintains an "observable" relationship with the environment that is called
Behavior.

1.9.5 Intelligence
The study of Kline [31] highlighted that the concept of intelligence is not, at the moment
very clearly defined, it does not seem to be a property of a system but a characteristic that can be
attributed to it by observing its behavior in the environment. It can be related to the capacity of
self-sufficiency. The behavior of an agent is intelligent if it can be justified as oriented towards
achieving specific objectives in its environment.

In Unified Theories of Cognition, Anderson [32] defines intelligence as: "the degree to
which a system approaches to make use of all available knowledge to face the solution of a
problem". Similarly in another study [33] it was proposed that the cognitive architecture of an
intelligent system must adapt optimally to the problems that the environment provides. For Xu and
Griffiths [34], a rational analysis takes into account the information available in the environment,
the objectives of the agent, and some criteria about the computational cost and produces the
optimal behavior function. This function must be tested in practice and adjusted to optimize its
performance.

1.9.6 Knowledge
According to Habermas [35] knowledge is defined as the information located in a context.
It is the information that is ready to be used in problem solving, to perceive and to learn.

1.10 Characteristics of a cognitive architecture


Any realistic model of the mind has to face at least several difficult demands: (1) to be, at
the same time, structured and flexible, (2) to face conditions of partial degradation of the internal
(organism) and external environment, (3) face multiple and simultaneous demands, (4) assign
quick meanings, without the need to resort to exhaustive algorithms, (5) complete conceptual,
perceptual patterns, etc., without having all the information logically necessary to do so, (6)
accessing their knowledge in different ways, and have memory systems in which they can be
recovered by virtue of their contents, and (7) adaptively modified according to previous
experiences.
BIOLOGICALLY INSPIRED REACTIVE PLANNER 17

A cognitive architecture is a broad framework in which different types of information


organizations have a place and where that information is subject to modifications in its content and
structure. Laird, Kinkade, Mohan, and Xu, [36] establishes some premises in which they points
out the need to differentiate the architecture and the information and the levels of the organization
that can be established within it. It is not possible to know the algorithm by which a behavior is
performed without knowing the primitive functions of the architecture. An architecture is needed
to predict the 'cognitive capacity' of an agent.

Bryson [37] argued that the architecture must be cognitively impenetrable, because
otherwise, there cannot be an algorithmic process operating on a representation to produce
behavior, because the internal structure of support can change. Rapp and Fischer-Baum [38]
suggested that cognitive architecture does not have to be something understandable and it is not
modifiable by the agent itself, it is the structural framework where cognitive processes are
developed, and the clear separation between the form and content of information is one of the
characteristics to be solved in a good model of cognitive architecture.

The idea is to associate architecture with the definition of certain coding and processing
mechanisms that will be concrete and predefined for the structure of each specific cognitive
architecture. These structural resources will support the contents. In addition to a syntactic
capability, the processing systems whether they are artificial or human, they need a certain amount
of knowledge specific problems to face a complex problem, 'semantically rich'.
BIOLOGICALLY INSPIRED REACTIVE PLANNER 18

Figure 4: Framework of Cognitive Architecture

1.11 Controlled Architectures


The control architectures can be computationally complex or simple, each with its pros and
cons, and depending on the desired application. Each of the control architectures allows the robots
or agents, in different degrees, to perform autonomously or semi-autonomously a variety of tasks
in different domains and applications. The selection of a control architecture depends mainly on
the intrinsic properties of the problem, the type of task desired, the required optimality and the
information available. In addition, in the physical implementation of the robot or agent, it can be
closely related to the restrictions of hardware and software, the morphology of the robot and the
interaction with the environment.
BIOLOGICALLY INSPIRED REACTIVE PLANNER 19

In this work the benefits of behavioral architecture have been implemented and shown, but
the other paradigms are also useful in various applications. For example, deliberative systems
provide reasoning and planning that is given by a precise representation of the environment. This
implies that the environment does not change; therefore, these systems are ideal for structured and
highly predictable environments, especially in situations where the agent repeatedly performs a
certain task.

On the other hand, reactive systems produce satisfactory results when dealing with
changing environments and where a response or immediate reaction is needed in the performance
of the task, such as the avoidance of obstacles. Therefore, changing (stochastic) environments are
best handled with reactive systems. In addition, reactive controllers are very powerful in
environments and tasks that have been characterized in advance and where the computational load
is not excessive. However, when you need the representation of the world and knowledge of past
and future actions, you can add other modules that contain planning, learning or memory.

Similarly, Stephens, Huang, Nemoto and Munro [39] have also supported hybrid systems
by stating that they have proven to be a good choice for environments that need internal models
and require long-term planning and that, in real time, the demands of the system are sufficiently
independent of reasoning of high level using specific languages and operating systems that can
respond quickly in dynamic environments.

Likewise, Arrichiello, Chiaverini, Indiveri and Pedone [40] stated that behavioral control
can be used as an alternative to traditional search methods such as hash tables, sequential search,
binary search, or the methods used by classical Artificial Intelligence such as search trees, method
of generation and testing, heuristic methods. In certain applications, these methods are not flexible
enough to respond successfully to different situations; good results are not obtained when the data
are inaccurate, incomplete or if more versatile searches are sought (for example using synonyms,
or through a user query). Some of the search methods, cited above, need a memory representation
of the search space and have a high computational load.

1.12 Behavioral Design


According to the study of Reimann et al. [41] two basic hierarchical behaviors were
considered, evasion of obstacles and monitoring of the right wall, the first with a higher level of
priority than the second. In the presence of frontal obstacles at distances less than or equal to 50
BIOLOGICALLY INSPIRED REACTIVE PLANNER 20

cm (safety distance), the obstacle avoidance behavior is activated, inhibiting the tracking of walls
and assuming control of the effectors of the robot until it is away from the collision path.

Once the conditions of deactivation of obstacle avoidance have been determined, the wall
tracking behavior is activated again and, through the control of the effectors, it will seek to reach
and maintain a reference distance of 50 cm between the robot and the objects located at its right
side (walls, columns, furniture, etc.)

Figure 5: Reactive Model

Only one behavior can be active at a given time and its activation or deactivation will only
depend on the state of the environment captured by the robot's sensors. The avoidance of obstacles
corresponds to a pure reactive behavior that is activated once the front sensor S3 detects an obstacle
at a distance less than or equal to 50 cm. The control module processes the measurement and
generates a sequence of actions to evade the obstacle, thus avoiding a collision.

These actions are: (1) inhibits the tracking of the right wall and stops both motors, (2)
activates the right motor until counting eight pulses of the encoder coupled to the right wheel, in
this way it is possible to change the orientation of the robot by approximately 90 degrees in anti-
clockwise direction and, (3) if no new obstacles are detected at distances less than or equal to 50
cm, the wall follow-up is reactivated, otherwise this same sequence is repeated.
BIOLOGICALLY INSPIRED REACTIVE PLANNER 21

1.13 Natural Selection in Ecology


Many investigations in evolutionary ecology quantify the magnitude and sign of
phenotypic selection. And in fact it is considered as natural selection without more. In this way,
natural selection can be defined concisely as a fitness difference between phenotypically different
biological entities. The variation between individuals in relative fitness is called opportunity for
selection. It can be measured as the standard deviation, the variance, the coefficient of variation,
etc. The possibilities of selection depend to a great extent on this variation in fitness.

The selective environment is a fundamental concept in evolutionary ecology: is defined as


e l set of those factors of ecological environment that determine the relative reproductive output
of individuals in the population. The environmental factor that is causing the trait-fitness
relationship that has been described in the definition of selection is known as selective pressure
[42].

Natural selection can be classified according to the effect it has on the statistical descriptors
of the phenotypic characters. If the selection acts on the average value of the characters, the
selection is said to be directional. If the selection works by decreasing the variance but keeping
the average of the value of the traits fixed, the selection is said to be stabilizing. If the selection
works by increasing the variance but without altering the average value of the traits, the selection
is said to be disruptive. The selection can also act on the covariance of the phenotypic traits. In
this case it is said that the selection is correlational. Finally, the selection can affect a trait not
directly but due to the existing correlation between this trait and another one that is being selected.
This type of selection is called indirect selection [43].

1.14 The second theorem of natural selection


Mark, Marion and Hoffman [44] stated that the rate of change by selection of the mean
value of a quantitative trait is proportional to the genetic covariance between trait and fitness. Most
of the evolutionary studies are aimed at determining the phenotypic evolution, which occurs on
features other than fitness, generically called z. We know that the evolution potential of a fitness
component is proportional to its variation. But if there is no variation in z, and all variants with
respect to fitness exhibit the same value of z, there is no possibility of evolution of z.
BIOLOGICALLY INSPIRED REACTIVE PLANNER 22

This relationship is called fitness function or fitness surface. For there to be a possibility
of evolution of z, it is necessary that this also varies between individuals, and that it varies in tune
with w. Picking up this idea in a formula, and moving the TFS to a trait z no fitness, we would
have that the rate of evolution is:

(1 / z) d z = V 2 z b wz

Where b wz is the slope of the relationship between w and z. It is known that:

b wz = Cov w, z / V z [Cov (w, z) = 1 / n Σ (wi-w) (zi-z)]

So:

(1 / z) d z = Cov w, z

In evolutionary ecology this theorem is even more important than TFS, since it offers us a
simple way to quantify the intensity of the selection on a character.

1.15 Exclusive selection of behaviors – Utility function method


In some systems, behaviors that are considered mutually exclusive are defined, in the sense
that they act on the same motor and actuator elements generating incompatible orders for them.
The need to establish in some way which of them is going to be selected at each moment is evident.
This problem has been widely studied in the field of Ethology, from which we can obtain
interesting conclusions for the design of robotic architectures based on behavior.

For example, an animal that must decide between 3 incompatible behaviors of drink, eat,
or runaway is influenced by several factors that could drive the execution of some of them over
the rest, some intrinsic, such as energy levels of the animal, and others dependent on the
environment, such as the presence of predators, or food: the magnitude of these stimuli would
increase or reduce the predisposition to the execution of a certain behavior [45].

A fundamental aspect is that the selection criteria must have a learning component based
on experience, since the behavior will be executed within a dynamic and changing environment.
In a complete architecture at all levels, a selection hierarchy would be established at different levels
of abstraction, so that the selection of each of the levels triggers a new selection mechanism at
lower levels, and so on. The lower levels would be related to the direct action on the motor and
actuator elements of the system [46].
BIOLOGICALLY INSPIRED REACTIVE PLANNER 23

In a complete architecture at all levels, a selection hierarchy would be established at


different levels of abstraction, so that the selection of each of the levels triggers a new selection
mechanism at lower levels, and so on. The lower levels would be related to the direct action on the
motor and actuator elements of the system.

1.16 Subsumption Architecture


The subsumption architecture, uses a fixed priority scheme in which the behavior of higher
layers can inhibit the responses of the lower layers, replacing them with their own. The most
complex aspects in this architecture are related to the establishment of priorities. The selection
criteria are usually associated with state machines and decision rules, and the appearance of
activating signals (releasers) of possible behaviors involved.

Figure 6: Subsumption Architecture

There are also selection architectures in which all competing behaviors are interconnected
in order to send reciprocal inhibitory signals, in addition to excitatory signals with shared
resources. Thus, an increase in the activity of a behavior increases the inhibition of the rest of the
behaviors and also the inhibitory action of them until, finally, some of the behaviors reach control
over the system. In this case the difficulty lies in defining the power of the excitatory and inhibitory
signals, and studying their dynamics of variation by their interaction [47].
BIOLOGICALLY INSPIRED REACTIVE PLANNER 24

As an additional drawback, the architecture has scalability problems since the inclusion of
new behaviors in the system requires the adjustment of as many additional inhibitory connections
as existing behaviors in the system and, possibly, a re-balancing of the dynamics of interaction
between the same. It should be noted that, although this type of distributed architecture may allow
the selection of behaviors in a "winner-take all" style, it would also be possible to take advantage
of the dynamics associated with constant changes due to feedback and variation in the stimuli of
entry to obtain emerging behaviors with mechanisms similar to those of organization of behaviors
by combination [48].

In contrast to the previous option, other architectures use centralized selection mechanisms
in which a central module is connected to all the behavior of the system reducing the number of
total connections, which favors the scalability and modularity of the system. The central module
would act as referee and selector of the winning behavior, and should receive from the rest of the
behaviors the adequate information that would allow it to establish a decision criterion. This
information would only be related to the coordination aspects of the behavior of the system, thus
facilitating the design of the control element. As a variant, the central arbitration module receives
global information from the system, as well as from the behaviors, and makes the selection using
ambiguous logic rules [49].

According to Gurney, Prescott, and Redgrave [50] it is also possible to combine some of
the architectures. In aspects of distributed and centralized selection architectures are combined, to
explain the functioning of the basal ganglia as an important component in the arbitration or
selection of behaviors in vertebrate animals. Dursun and Karsak [51] proposed a conceptual model
of the selection system proposed for the basal ganglia, for an example of control of motor resources
shared by some behaviors or command systems, called "channels" in the image.

These channels generates excitatory signals and inhibitors towards the motor resources and
the central arbitration element (i.e., the basal ganglia), which in turn generates new signals of
inhibition or excitation towards the channels so that, finally, a single channel or behavior is
selected for the execution of motor actions. This same scheme could be adapted to methods of
emergent combination of behaviors eliminating the final step of inhibition of the "losers" and
allowing a contribution of each of them proportional to the level of the signal sent from the central
element.
BIOLOGICALLY INSPIRED REACTIVE PLANNER 25

1.17 Expert Systems


The AI design philosophies aspire, to a great extent, the development of a model or
framework of general intelligence, that allows to emulate the way in which human beings solve
problems and face the situations found in the day to day of his life. There are, however, other
currents in the field of research in AI, which consider this objective too ambitious and that defend,
on the other hand, the development of procedures and techniques focused on the resolution of
problems circumscribed to very specific areas of intelligence [52].

This set of methods and algorithms are included in what is known as Weak or Reduced
Artificial Intelligence. And, although they do not coincide in the vision of building global
architectures that allow us to develop AI similar to human ones, they are often part of the most
elaborate and complete architectures, in many cases as a method of implementing any of those
stages.

The techniques known collectively as Expert Systems, try to incorporate into a computer
the human experience in a certain field, looking for the artificial system to be able to fulfill its task
at the same level as the human expert. An important problem to achieve this objective is that human
experts often do not have clear rules used to make their decisions, basing many of them on general
rules according to their experience (heuristics), and even in intuitions. Some early examples of this
type of system were DENDRAL, which is considered as the first expert system, and applied to
chemistry; MYCIN, in the field of medicine; and PROSPECTOR, in the field of geology [53].

All of them were adding new advances in the area, especially related to the methods of
capturing, analyzing, and expressing information and rules to make their decisions: this is what is
currently known as Knowledge Engineering. Gradually, aspects such as the separation of the rules
and the reasoning mechanism in different blocks were added; or the uncertainty, in many cases, of
knowledge, and how to deal with this problem. Among the characteristics common to this type of
systems are related to its application in a domain or very specific area; the use of knowledge based
on experience; the ability to explain the decisions taken; and symbolic reasoning, which allows to
represent different types of knowledge, such as concepts, evidence, or rules, using qualitative
rather than numerical data [54].
BIOLOGICALLY INSPIRED REACTIVE PLANNER 26

References

[1]. Wortham, Robert H., Swen E. Gaudl, and Joanna J. Bryson. "Instinct: A biologically inspired
reactive planner for embedded environments." In Proceedings of ICAPS 2016 PlanRob
Workshop, London, UK. 2016.
[2]. Wortham, Robert H., and Joanna J. Bryson. "A Role for Action Selection in Consciousness:
An Investigation of a Second-Order Darwinian Mind." EUCognition. 2016.
[3]. Mese, Erkan, and David A. Torrey. "An approach for sensorless position estimation for
switched reluctance motors using artifical neural networks." IEEE Transactions on power
electronics 17, no. 1 (2002): 66-75.
[4]. Karouni, Ali, Bassam Daya, and Samia Bahlak. "Offline signature recognition using neural
networks approach." Procedia Computer Science 3 (2011): 155-161.
[5]. Low, Kian Hsiang, Wee Kheng Leow, and Marcelo H. Ang. "Integrated planning and control
of mobile robot with self-organizing neural network." In Proceedings 2002 IEEE International
Conference on Robotics and Automation (Cat. No. 02CH37292), vol. 4, pp. 3870-3875. IEEE,
2002.
[6]. Shanmuganathan, Subana. "Artificial neural network modelling: An introduction." In Artificial
neural network modelling, pp. 1-14. Springer, Cham, 2016.
[7]. Murphy, Kevin P. Machine learning: a probabilistic perspective. MIT press, 2012.
[8]. Mohri, Mehryar, Afshin Rostamizadeh, and Ameet Talwalkar. Foundations of machine
learning. MIT press, 2018.
[9]. Prescott, Tony J., Joanna J. Bryson, and Anil K. Seth. "Introduction. Modelling natural action
selection." (2007): 1521-1529.
[10]. Snoek, Jasper, Hugo Larochelle, and Ryan P. Adams. "Practical bayesian optimization of
machine learning algorithms." In Advances in neural information processing systems, pp.
2951-2959. 2012.
[11]. Brooks, Rodney A. "Intelligence without reason." (1991).
[12]. Gaudl, Swen E., and Joanna J. Bryson. "Extended ramp goal module: Low-cost behaviour
arbitration for real-time controllers based on biological models of dopamine cells." In 2014
IEEE Conference on Computational Intelligence and Games, pp. 1-8. IEEE, 2014.
BIOLOGICALLY INSPIRED REACTIVE PLANNER 27

[13]. Trivedi, Deepak, Christopher D. Rahn, William M. Kier, and Ian D. Walker. "Soft robotics:
Biological inspiration, state of the art, and future research." Applied bionics and
biomechanics 5, no. 3 (2008): 99-117.
[14]. Winsberg, Eric. Science in the age of computer simulation. University of Chicago Press,
2010.
[15]. Brom, Cyril, Jakub Gemrot, Michal Bida, Ondrej Burkert, Sam J. Partington, and J. Bryson.
"Posh tools for game agent development by students and non-programmers." In The Nineth
International Computer Games Conference: AI, Mobile, Educational and Serious Games,
pp. 126-133. University of Bath, 2006.
[16]. Drumwright, Evan, John Hsu, Nathan Koenig, and Dylan Shell. "Extending open dynamics
engine for robotics simulation." In International Conference on Simulation, Modeling, and
Programming for Autonomous Robots, pp. 38-50. Springer, Berlin, Heidelberg, 2010.
[17]. Breazeal, Cynthia, and Brian Scassellati. "Robots that imitate humans." Trends in cognitive
sciences 6, no. 11 (2002): 481-487.
[18]. Nilsson, Nils J. Shakey the robot. SRI INTERNATIONAL MENLO PARK CA, 1984.
[19]. Wortham, Robert H., Andreas Theodorou, and Joanna J. Bryson. "What does the robot think?
Transparency as a fundamental design requirement for intelligent systems." In Ijcai-2016
ethics for artificial intelligence workshop. 2016.
[20]. Somerville, Leah H., and B. J. Casey. "Developmental neurobiology of cognitive control and
motivational systems." Current opinion in neurobiology 20, no. 2 (2010): 236-241.
[21]. Seth, Anil K. "The ecology of action selection: Insights from artificial life." Philosophical
Transactions of the Royal Society B: Biological Sciences 362, no. 1485 (2007): 1545-1558.
[22]. Varela, Francisco J., Evan Thompson, and Eleanor Rosch. The embodied mind: Cognitive
science and human experience. MIT press, 2017.
[23]. Clark, Andy. "Whatever next? Predictive brains, situated agents, and the future of cognitive
science." Behavioral and brain sciences 36, no. 3 (2013): 181-204.
[24]. Volos, Ch K., Ioannis M. Kyprianidis, and Ioannis N. Stouboulos. "A chaotic path planning
generator for autonomous mobile robots." Robotics and Autonomous Systems 60, no. 4
(2012): 651-656.
[25]. Williams, George C. Adaptation and natural selection: A critique of some current
evolutionary thought. Vol. 61. Princeton university press, 2018.
BIOLOGICALLY INSPIRED REACTIVE PLANNER 28

[26]. Farinelli, Alessandro, Manuele Bicego, Sarvapali Ramchurn, and Mauro Zucchelli. "C-link:
A hierarchical clustering approach to large-scale near-optimal coalition formation."
In Twenty-Third International Joint Conference on Artificial Intelligence. 2013.
[27]. Gil-García, Reynaldo, José M. Badía-Contelles, and Aurora Pons-Porrata. "Dynamic
hierarchical compact clustering algorithm." In Iberoamerican Congress on Pattern
Recognition, pp. 302-310. Springer, Berlin, Heidelberg, 2005.
[28]. Gil-García, Reynaldo, and Aurora Pons-Porrata. "Dynamic hierarchical algorithms for
document clustering." Pattern Recognition Letters 31, no. 6 (2010): 469-477.
[29]. Sowa, John F., ed. Principles of semantic networks: Explorations in the representation of
knowledge. Morgan Kaufmann, 2014.
[30]. Shapiro, Stuart C., and Jonathan P. Bona. "The GLAIR cognitive architecture." International
Journal of Machine Consciousness 2, no. 02 (2010): 307-332.
[31]. Kline, Paul. Intelligence: The psychometric view. Routledge, 2013.
[32]. Anderson, Norman H. A functional theory of cognition. Psychology Press, 2014.
[33]. Salvucci, Dario D. "Modeling driver behavior in a cognitive architecture." Human
factors 48, no. 2 (2006): 362-380.
[34]. Xu, Jing, and Thomas L. Griffiths. "A rational analysis of the effects of memory biases on
serial reproduction." Cognitive psychology 60, no. 2 (2010): 107-126.
[35]. Habermas, Jürgen. Knowledge and human interests. John Wiley & Sons, 2015.
[36]. Laird, John Edwin, Keegan R. Kinkade, Shiwali Mohan, and Joseph Z. Xu. "Cognitive
robotics using the soar cognitive architecture." In Workshops at the Twenty-Sixth AAAI
Conference on Artificial Intelligence. 2012.
[37]. Bryson, Joanna Joy. "The study of sequential and hierarchical organisation of behaviour via
artificial mechanisms of action selection." PhD diss., University of Edinburgh, 2000.
[38]. Rapp, Brenda, and Simon Fischer-Baum. "Uncovering the cognitive architecture of
spelling." In The handbook of adult language disorders, pp. 75-102. Psychology Press, 2015.
[39]. Stephens, Ashley M., Jingjing Huang, Kae Nemoto, and William J. Munro. "Hybrid-system
approach to fault-tolerant quantum communication." Physical Review A 87, no. 5 (2013):
052333.
BIOLOGICALLY INSPIRED REACTIVE PLANNER 29

[40]. Arrichiello, Filippo, Stefano Chiaverini, Giovanni Indiveri, and Paola Pedone. "The null-
space-based behavioral control for mobile robots with velocity actuator saturations." The
International Journal of Robotics Research 29, no. 10 (2010): 1317-1337.
[41]. Reimann, Martin, Judith Zaichkowsky, Carolin Neuhaus, Thomas Bender, and Bernd Weber.
"Aesthetic package design: A behavioral, neural, and psychological investigation." Journal
of Consumer Psychology 20, no. 4 (2010): 431-441.
[42]. Smith, Eric Alden, and Bruce Winterhalder. "Natural selection and decision-making: Some
fundamental principles." In Evolutionary ecology and human behavior, pp. 25-60.
Routledge, 2017.
[43]. Schoener, Thomas W. "The newest synthesis: understanding the interplay of evolutionary
and ecological dynamics." science 331, no. 6016 (2011): 426-429.
[44]. Mark, Justin T., Brian B. Marion, and Donald D. Hoffman. "Natural selection and veridical
perceptions." Journal of Theoretical Biology 266, no. 4 (2010): 504-515.
[45]. Wei, Ermin, Asuman Ozdaglar, and Ali Jadbabaie. "A distributed Newton method for
network utility maximization—Part II: Convergence." IEEE Transactions on Automatic
Control 58, no. 9 (2013): 2176-2188.
[46]. Das, Rajarshi, Jeffrey Owen Kephart, Gerald James Tesauro, and William Edward Walsh.
"Method and apparatus for utility-based dynamic resource allocation in a distributed
computing system." U.S. Patent 8,352,951, issued January 8, 2013.
[47]. Colledanchise, Michele, Richard M. Murray, and Petter Ögren. "Synthesis of correct-by-
construction behavior trees." In 2017 IEEE/RSJ International Conference on Intelligent
Robots and Systems (IROS), pp. 6039-6046. IEEE, 2017.
[48]. Oland, Espen, Tom Stian Andersen, and Raymond Kristiansen. "Subsumption architecture
applied to flight control using composite rotations." Automatica 69 (2016): 195-200.
[49]. Nagata, F., Otsuka, A., Watanabe, K., & Habib, M. K. (2013). Multiple mobile robots system
with network-based subsumption architecture. International Journal of Mechatronics and
Manufacturing Systems, 6(1), 57-71
[50]. Gurney, Kevin N., Tony J. Prescott, and Peter Redgrave. "The basal ganglia viewed as an
action selection device." In International Conference on Artificial Neural Networks, pp.
1033-1038. Springer, London, 1998.
BIOLOGICALLY INSPIRED REACTIVE PLANNER 30

[51]. Dursun, Mehtap, and E. Ertugrul Karsak. "A fuzzy MCDM approach for personnel
selection." Expert Systems with applications 37, no. 6 (2010): 4324-4330.
[52]. Abu-Naser, S. S., H. El-Hissi, M. Abu-Rass, and N. El-Khozondar. "An expert system for
endocrine diagnosis and treatments using JESS." Journal of Artificial Intelligence 3, no. 4
(2010): 239-251.
[53]. Hadavandi, Esmaeil, Hassan Shavandi, and Arash Ghanbari. "Integration of genetic fuzzy
systems and artificial neural networks for stock price forecasting." Knowledge-Based
Systems 23, no. 8 (2010): 800-808.
[54]. Russell, Stuart J., and Peter Norvig. Artificial intelligence: a modern approach. Malaysia;
Pearson Education Limited,, 2016.

You might also like