Abstract
Neural networks have been defined as mathematical constructs that have the ability to emulate the numerous processes individuals utilize to recognize patterns, learn tasks and solve problems. For a long time, Neural networks have been characterized in terms of their number and types of connections specifically between the person and processing elements known as neurons and the learning rules utilized when data is presented to the network. Therefore, the basis of this paper is to critically look at the concept of neural networks and the way they are employed in the decision-making process.
As a result, the paper will answer questions such as what are Neural networks, how do they work, how useful are neural networks in solving problems, and what current evidence exists to give an understanding of the scientific and technical advancement of Neural networks? The paper, in providing answers to these questions, will heavily rely on and utilize theories of March, Simon and Rappaport from the Breath component while at the same time utilizing the theory of integration of neural networks. Hence, the overall aim of the discussion will be to find out the advantages of neural networks in decision-making.
Introduction
“A neural network is a powerful data modeling tool that can capture and represent complex input/output relationships” (Neural Topics, 2010, p.1). In creating and developing neural network technology, the overriding motivation originated from the aspiration to grow an artificial system that possesses the ability to execute ‘intelligent’ responsibilities that are in one way similar to those done by the human brain. Further, it was identified that neural networks have the “ability to represent linear and non-linear relationships equally and their capacity to learn these relationships straight from the data modeled” (Neural Topics, 2010, p.1).
In essence, neural networks resemble brains whereby they have been perceived to be good at analyzing information and recognizing patterns that are complex to delineate accurately. In addition, Neural networks are trained where thousands of examples are used and a learning algorithm which has the potential to change the strength of the connections in the network with objective of giving it an appropriate output value which in large measures it depends on the input values (Neural Topics, 2010, p.1).
Scientists have found out that the human brain is an exceptionally exciting information processor, though it is incapable of working faster than the ordinary computer. Many of these researchers in artificial intelligence have investigated the organization of the brain as a model for constructing intelligent apparatus. According to George Luger, the constructed neural models of intelligence generally focus on the ability of the brain to adjust to the world in which it is placed through the process of modifying the associations between the individual neurons. In addition, “rather than representing knowledge in explicit logical sentences, they capture it implicitly, as a property of patterns of relationships” (Neural Topics, 2010, p.1).
Moreover, according to InformationWeek article of 2005, neural networks unlike other discoveries in artificial intelligence, have limited ability to manipulate and relate symbols about concepts in the world, instead, neural networks function in accordance to the lists of numbers that represent numerous problems and their likely solutions (cited in Neural Topics, 2010, p.1). At the same time, “the artificial neurons have the capacity to learn relationships based on a training set of solutions and eventually become stacked into ‘layers’ so that output of one neural network could form the input of another” (Neural Topics, 2010, p.1).
Decision-making
In recent times, there has been huge evidence indicating that interest has grown specifically in developing capable decision systems within the field of Artificial Intelligence (AI). Decision-making can be defined as the process that involves the ability to choose some course of action among the various alternatives available (Zopounidis and Pardalos, 1998). In addition, decision-making is affected by numerous factors that may include the volume of data involved and the tradeoffs that key people involved in decision-making have to face in real life practical situations.
Observing the process, one discovers that in almost all circumstances decision problems possess multiple and contradictory criteria for judging the possible alternatives. As a fact, the major objective of the decision-maker is to fulfill his or her contradictory goals while at the same time possibly satisfying the constraints of the system (Zopounidis and Pardalos, 1998). As a result, the scientific community in the past decades has been involved in intensive discovery and development of systems that can perform decision making instead of humans. Interest has led scientists to develop a learning system known as neural networks.
For instance, the conviction in the field is that a class of problems involving decision-making relates to the interpretation of data based on a specific set of rules, and in such a case, the neural network learns the rules governing the decision-making by experiments. Further, in another class of problems, the simulation of systems can facilitate decision-making with complicated nonlinear dynamics by neural networks (Zopounidis and Pardalos, 1998).
Theories of Neural Networks
One of the vital existing problems in Artificial Intelligence has been identified as, how to enable computers to draw logical conclusions from real and existing facts which are largely known as the process of reasoning (Levy, 2006). For instance, to prove a theorem in logic or to solve a particular problem that needs the use of logic, a proof is always needed, whereby a proof constitutes a sequence of expressions where every expression in the sequence is either one of the already accepted theorems or axioms or an expression that can be found from one or two of the already accepted theorems or axioms whereby one needs to apply the rules of inference (Levy, 2006).
The Logic Theory Machine
Allen Newell, Herbert Simon, and John Show developed this theory in 1956; the theory outlines the proofs to theorems in a form of logic known as the propositional calculus. Logic theory machine uses the axioms and inference rules that were proposed by Russell and Whitehead, which implies that in order to function on mathematical theorems that have been approved, it is important to generate new theorems.
For instance, the theory postulate that, when there is an expression to be proved, five-set of axioms are established, and using the theory, three inference rules are used repeatedly until they produce the desired mathematical expression which in this case is the one to be proved. As such, a computer can be given an infinite amount of time and memory, while the theory can be used to prove any provable theorem and solve any solvable problem (Levy, 2006).
In this case, the computer constructs a program, which eventually constructs subsequent possible proofs in a systematic manner but in a random order, checking each time to eliminate duplicate attempts and to see if the target theorem has been proved. To make the theory more efficient, radical improvements in the possible proofs of theorems and solutions to problems have been made through developed methods of substitution, detachment and chaining (Levy, 2006).
Valence Instrumentality-Expectancy (VIE) Theory
This theory postulate that individual’s actions and choices are legitimately associated to the preferences and sentimental reactions they possess for specific expectations (valences), in addition to the individual’s beliefs about whether particular actions results to specific outcomes or performance level (expectancies) and the individual’s perception of the relationship existing between primary and secondary outcomes (instrumentalities). In most occasions, people are believed to evaluate their valences, instrumentalities, and expectancies knowingly when making decisions. Further, the belief is that the same people act in a subjectively optimal way especially when the factors have been considered (Lord, Hanges, and Godfrey, 2003).
This theory and ideas influenced decision-making research activities that influence some aspects of organization such as leadership, compensation and turnover (Lord, Hanges, and Godfrey, 2003). Unfortunately, the discovery on this model indicated that it is inconsistent with the informational and processing demands of other optimization models. This has led to improvement of the model where the belief is that the manifested inconsistent can be resolved by assuming that computations implied by optimization models are generally performed automatically by neural networks. The realization in accordance to this conviction is that neural networks are apposite computational devices that are effective in implementing procedures of VIE (Lord, Hanges and Godfrey, 2003).
Neural Networks and Decision-making
Paul Thagard and Brandon Wagar note that some individuals have the tendency to make decisions by flipping a coin (Wagar and Thagard, 2004 cited in Thagard and Kroon, 2006, p.87). To such individuals, the issue is not to make the decision indicated by the flip but to see how they feel about the choice that the coin flips (Wagar and Thagard, 2004 cited in Thagard and Kroon, 2006). To the authors, the act of flipping a coin has evolved as an effective way to discover the individual’s emotional reactions to numerous alternatives and point out the emotional weight they connect to such decisions.
As such, there exists wide consensus in the field of cognitive science that emotions constitute an integral part of decision-making (Wagar and Thagard, 2004 cited in Thagard and Kroon, 2006). Today, there are artificial neural network models of cognitive affective processing that make use of simplified neurons and in the process, they have neglected neuroanatomical information about how different parts of the brain contribute to decision making.
The authors present a new computational model known as GAGE that has more neural characteristics where it organizes neurons into anatomically recognized groups that correspond to vital brain areas including ventromedial prefrontal cortex (VMPFC), the hippocampus and the amygdale (Wagar and Thagard, 2004 cited in Thagard and Kroon, 2006). The authors identified further the nucleus accumbens (NAcc) as another vital component that functions to incorporate “cognitive information from the ventromedial prefrontal cortex and the hippocampus with emotional information from the amygdala” (Wagar and Thagard, 2004 cited in Thagard and Kroon, 2006).
This model was implemented in a computer program that successfully simulates two types of cognitive-affective integration: people’s performance on the Iowa Gambling Task and the integration of physiological arousal and cognition in determining emotional states (Wagar and Thagard, 2004 cited in Thagard and Kroon, 2006).
The VMPFC is seen to be involved in the production of somatic markers, which in essence are the feelings, or emotional reactions that are largely connected through experience with the likelihood of long-term outcomes of specific responses to a particular given situation. Accordingly, the hypothesis postulates that sensory representations of particular response to the current situation activate knowledge tie in previous emotional experience and the resultant concealed emotional signals operate as biases that have the ability to influence the mechanisms responsible for higher-level cognitive processes.
These somatic markers have been useful in decision making processes where they narrow the number of feasible behavioral alternatives, while at the same time allowing the organism to reason according to the long-term predicted outcomes of its actions (Wagar and Thagard, 2004 cited in Thagard and Kroon, 2006).
On his part, Walter Bischof establishes that decision making originates from a given set of patterns each described by a set of continuous attributes, and in such case, linear discriminant functions are employed to separate the patterns by a linear decision boundary (Bischof, 2004). In circumstances where multiple pattern classes exist, the feature space must be divided into multiple partitions where each is associated with a single class; thus, using discriminant functions, this is achieved in several ways (Bischof, 2004).
In current neural networks, specifically multi-layered perceptions have become accepted for pattern learning and as such, neural networks can be compared in their performance to discriminant functions (Bischof, 2004). In this pattern learning, patterns are represented by specific characteristic features. The process consists of finding a partitioning of the feature space such that each partition corresponds to a single pattern class, while pattern classification is based on combinations of discriminant functions (Bischof, 2004). Further, the author notes how the decision tree has become popular in decision-making process.
The decision tree method aims at ordering the patterns in a tree where each tree node specifies a test of some attribute and the process of moving from the tree to a leaf represents a hierarchically ordered sequence of decisions for classifying the patterns (Bischof, 2004).
Application: the Use of Complementary Decision Learning System (CDMS)
Neural network shows characteristics such as human-like decision-making process and possesses steps that resemble human decision-making; thus, CDMS has been developed base on fuzzy neural networks (Jain and Lim, 2010). First, it can be seen that human decision-making process involves lateral inhibition between reward and punishment stimuli and isolation of reward and punishment knowledge; hence, CDMS is implemented with complimentary learning. CDMS exhibit similar rules as those of neurobiological decision-making where it allows the reward rules to fire and inhibit the punishment rules.
However, when the punishment-related stimulus is presented, rules associated to punishment will be activated and in tandem rules related to rewards are inhibited (Jain and Lim, 2010). In functioning, once the CDMS has been decided, the firing strength of the winning rule is utilized to find out the action to be taken. After the CDMS has outputted the action, the reward system evaluates whether the decision made brings about reward or punishment, in the same measure the reward system learns to predict the outcome of CDMS decision.
In order to function effectively, the CDMS is equipped with human-like reasoning process and mapped onto a hypothetico-deductive fuzzy scheme referred to as Analogical Approximate Reasoning Schemes (AARS). This is one of the ways physicians make decisions and when the mapping is done, “it not only formalizes the operation of CDMS, but also equips CDMS with decision-making process analogous especially to the physician” (Jain and Lim, 2010). The CDMS hence demonstrate a reasoning process similar to that of human.
Advantages of Neural Networks in Decision-making
Specifically, employing artificial neural networks in decision-making has been widely believed to have the potential to solve the numerous forecasting and decision modeling problems (Hill, Marquez, O’Connor and Nemus, 1993). For instance, neural networks are able to model any kind of parametric or non-parametric process at a fast rate and automatically alter the input data (Hill, Marquez, O’Connor and Remus, 1993). Second, the artificial neural networks have universal function approximators for even non-linear functions and they have the ability to estimate piece-wise approximations of functions.
Third, artificial neural networks have the ability to deal with intricate data in a more successful way when compared to regression analysis (Anagnostou, Remzi, and Djavan, 2003). In addition, neural networks have been described as the best alternatives to convectional techniques, which in most ways are limited by strict assumptions of normality, linearity, and variable independence. Therefore, neural networks have the ability to capture numerous kinds of relationships since they allow the person using them to swiftly and easily model phenomena which otherwise may have been extremely intricate or impossible to explain.
According to Chris Stergiou, neural networks are generally useful since they have ability to develop meaning from problematical data while at the same time being able to extract patterns and identify trends that appear too intricate to be noticed by human or other different computer methods. In summary, the author notes the following as the main advantages of neural networks: they largely demonstrate: “adaptive learning, self-organization, real time operation, and fault tolerance via redundant information coding” (Stergiou, n.d, p.1).
Conclusion
Neural network is a fast growing field and its relevance is being demonstrated in numerous fields. The major importance of the concept has widely being utilized in decision-making processes. For instance, today there is a refined computational intelligence technique that is aiding the process of decision making especially under uncertain conditions through numerous means such as coordinating data delivery, analyzing data trends, providing forecasts, ensuring data consistency, quantifying uncertainty, anticipating the user’s data needs, providing information to the user in the most appropriate forms and suggesting courses of actions.
In such a way, neural networks are graduating as techniques that have the ability to effectively describe complex behavior that is difficult to describe mathematically using the convectional methods. Nevertheless, with the discovery of some limitations of the technique, more advancement to improve the technique is necessary to ensure effectiveness and efficiency of the technique especially in relation to the outcomes.
References
Anagnostou, T., Remzi, M. and Djavan, B. (2003). Artificial Neural Networks for Decision-Making in Urologic Oncology. Austria, University of Vienna. Web.
Bischof, W. F. (2004). Visual Learning: An Overview. Canada, University of Alberta. Web.
Hill, T., Marquez, L., O’Connor, M. and Remus, W. (1993). Artificial Neural Network models for forecasting and decision-making. Web.
Jain, L. C., and Lim, C. P. (2010). Handbook on Decision Making: Techniques and Applications. Berlin, Springer. Web.
Levy, D. N. L. (2006). Robots unlimited: life in a virtual age. MA, AK Peters Ltd. Web.
Lord, R. G., Hanges, P. J. and Godfrey, E. G. (2003). Integrating neural networks into decision-making and motivational theory: Rethinking VIE theory. Canadian Psychology. Web.
Neural Topics. (2010). Neural Networks: Learning by Training a Network or Connectionist System. Web.
Stergiou, C. (N.d). What is a Neural Network? Web.
Wagar, B. and Thagard, P. (2006). Spiking Phineas Gage: A Neurocomputational Theory of Cognitive-Affective Integration in Decision Making in Thagard, P. and, Kroon, F. (2006). Hot thought: mechanisms and applications of emotional cognition. MA, MIT Press. Web.
Zopounidis, C. and Pardalos, P. M. (1998). Managing in uncertainty: theory and practice. MA, Kluwer Academic Publishers. Web.