**1. Introduction**

262 Bio-Inspired Computational Algorithms and Their Applications

[28] S. Kwong, Q. H. He, K. F. Man, K. S. Tang, and C. W. Chau, Parallel Genetic-based

[29] V. V. Ngoc, J. Whittington, and J. Devlin, Real-time Hardware Feature Extraction with

[30] S. A. Selouani and D. O'Shaughnessy, Robustness of Speech Recognition using Genetic

[31] L. R. Rabiner, Tutorial on Hidden Markov Models and Selected Applications in Speech Recognition, *Proceedings of the IEEE*, vol. 77, no. 2, pp. 257-286, 1989. [32] S. Okawa, E. Bocchieri, and A. Potamianos, Multi-Band Speech Recognition in Noisy

[33] A. de la Torre et al., Speech Recognition Under Noise Conditions: Compensation

[34] A. Álvarez et al., Application of Feature Subset Selection Based on Evolutionary

[35] G. E. Dahl, D. Yu, L. Deng, and A. Acero, Large Vocabulary Continuous Speech

[36] B. M. Ben-David et al., Effects of Agging and Noise on Real-Time Spoken Word

*Technology Conference*, Budapest, Hungary, 2001, pp. 1926-1931.

*Technologies*. Vienna, Austria: In-tech, 2011, cap. 2, pp. 29-54.

*ICASSP 1998*, Prague, Czech Republic, 1998, pp. 1-4.

Vienna, Austria: In-Tech, 2007, cap. 25, pp. 440-460.

*Hearing Research*, vol. 54, pp. 243-262, February 2011.

1998.

May 2007.

pp. 1-4.

Canada, 2004, pp. I-201-4.

Conversational Telephone Speech, in *IEEE Instrumentation and Measurement* 

Hybrid Pattern Matching Algorithm for Isolated Word Recognition, *International Journal of Pattern Recognition and Artificial Intelligence*, vol. 12, no. 4, pp. 573-594,

Embedded Signal Enhancement for Automatic Speech Recognition, in *Speech* 

Algorithms and Mel-cepstral Subspace Approach, in *IEEE International Conference on Acoustics, Speech, and Singal Processing 2004 ICASSP '04*, Montreal, Quebec,

Environments, in *International Conference on Acoustics, Speech and Signal Processing* 

Methods, in *Speech Recognition and Understanding*, M. Grimm and K. Kroschel, Eds.

Algorithms for Automatic Emotion Recognition in Speech, *Lecture Notes in Computer Science. Advances in Nonlinear Speech Processing,* vol. 4885, pp. 273-281,

Recognition with Context Dependent DBN-HMMS, in *International Conference on Acoustics, Speech and Signal Processing ICASSP 2011*, Prague, Czech Republic, 2011,

Recognition: Evidence from Eye Movements, *Journal of Speech, Language, and* 

Genetic algorithm is one of the successful optimization algorithm used in computing to find exact or approximate solutions for certain complex problems. This novel algorithm was first introduced by John Holland in 1975 (Holland, 1975). Besides Holland, many other researchers have also contributed to genetic algorithm (Davis, 1987; Davis, 1991; Grefenstte, 1986; Goldberg, 1989; Michalewicz, 1992). This is an algorithm that imitates the evolutionary process concept based on the Darwinian Theory which emphasizes on the law of "the survival of the fittest". This algorithm used techniques which are inspired from evolution biology such as inheritance, selection, crossover and mutation (Engelbrecht, 2002).

There are several important components in genetic algorithm which includes representation, fitness function, and selection operators (parent selection and survivor selection, crossover operator and mutation operator). Genetic algorithm starts by generating an initial population of individuals randomly. The individuals are represented as a set of parameter which is the solution to the problem domain. Normally, individuals are fixed length binary string. The individuals are then evaluated using fitness functions. The evaluation will give a fitness score to individuals indicating how well the solutions perform in the problem domain. The individuals that have been evaluated using the fitness function will be selected to be parents to produce offspring through the crossover and mutation operators. The genetic algorithms will repeat the above process except for the population initialization until the termination criteria is met. Fig. 1 shows the structure of a genetic algorithm.

GAs have been applied successfully in many applications including job shop scheduling (Uckun *et al.* 1993), the automated design of fuzzy logic controllers and systems (Karr 1991; Lee & Takagi, 1993), hardware-software co-design and VLSI design (Catania *et al.* 1997; Chandrasekharam *et al.* 1993). In this chapter, variations of genetic algorithms are applied in optimizing the bidding strategies for a dynamic online auctions environment.

Auction is defined as a bidding mechanism and is expressed by a set of auction rules that specify how the winner is determined and how much he or she has to pay (Wolfstetter, 2002). Jansen defines an online auction as an Internet-based version of a traditional auction (Jansen, 2003). In today's e-commerce market, online auction has acted as an important tool

Performance of Varying Genetic Algorithm Techniques in Online Auction 265

To address the shortcomings mentioned above, an autonomous agent was developed that can participate in multiple heterogeneous auctions. It is empowered with trading capabilities and it is able to make purchases autonomously (Anthony, 2003; Anthony & Jennings, 2003b). Two primary values that heavily influenced the bidding strategies of this agent are the *k* and *β*. These two values correspond to the polynomial function of the four bidding constraints, namely the remaining time left, the remaining auction left, the user's desire for bargain and the user's level of desperateness. Further details on the strategies will be discussed in Section 3. The *k* value ranges from 0 to 1 while the *β* value is from 0.005 to 1000. The possible combinations between these two values are endless and thus, the search space for the solution strategies is very large. Hence, genetic algorithms were used to find

This work is an extension of the solution above, which has been successfully employed to evolve effective bidding strategies for particular classes of environment. This work is to improve the existing bidding strategy through the optimization techniques. Three different variations of genetic algorithm techniques are used to evolve the bidding strategies in order to search for the nearly optimal bidding solution. The three techniques are parameter tuning, deterministic dynamic adaptation, and self-adaptation. Each of this method will be detailed in Section 4, 5 and 6. The remainder of the chapter is organized as follow. Section 2 discusses related work. The bidding strategy framework is discussed in Section 3. The parameter tuning experiment is discussed in Section 4. Section 5 and 6 discussed the deterministic adaptive experiment and self-adaptive experiment. A comparison between all the schemes is discussed in Section 7. Finally, the conclusion is

Genetic algorithm has shown to perform well in the complex system by which the old search algorithm has been solved. This is due to the nature of the algorithms that is able to discover optimal areas in a large search space with little priori information. Many researches in auctions have used genetic algorithm to design or enhance the auction's bidding strategies. The following section discusses works related to evolving bidding strategies.

An evolutionary approach was proposed by Babanov (2003) to study the interaction of strategic agents with the electronic marketplace. This work describes the agents' strategies based on different methodologies that employ incompatible rules in collecting information and reproduction. This work used the information collected from the evolutionary framework for economic studies as many researches have attempted to use evolutionary frameworks for economics studies (Nelson, 1995; Epstein & Axtell, 1996; Roth, 2002; Tesfatsion, 2002). This evolutionary approach allows the strategies to be heterogeneous rather than homogenous since only a particular evolutionary approach is applied. This work has shown that the heterogeneous strategies evolved from this framework can be used as a

ZIP, introduced by Cliff, is an artificial trading agent that uses simple machine learning to adapt and operate as buyers or sellers in online open-outcry auction market environments (Cliff, 1997). The market environments are similar to those used in Smith's (Smith, 1962)

the nearly optimal bidding strategy for a given auction environment.

discussed in Section 8.

**2. Related work** 

useful research data.

in the services for procuring goods and items either for commercialize purposed or for personal used. Online auctions have been reported as one of the most popular and effective ways of trading goods over the Internet (Bapna *et al.* 2001). Electronic devices, books, computer software, and hardware are among the thousands items sold in the online auctions every day. To date, there are 2557 auction houses that conduct online auctions as listed on the Internet (Internet Auction List, 2011). These auction houses conduct different types of auctions according to a variety of rules and protocols. eBay, as one of the largest auction house alone has more than 94 million registered users and had transacted more than USD 92 billion worth of goods during 2010 (eBay, 2010). These figures clearly show the importance of online auctions as an essential method for procuring goods in today's ecommerce market.


Fig. 1. The structure of a Genetic Algorithm

The auction environment is highly dynamic in nature. Since there are a large number of online auction sites that can be readily accessed, bidders are not constrained to participate in only one auction; they can bid across several alternative auctions for the same good simultaneously. As the number of auction increases, difficulties such as monitoring the process of auction, tracking bid and bidding in multiple auctions arise when the number of auctions increases. The user needs to monitor many auctions sites, pick the right auction to participate, and make the right bid in order to have the desired item. All of these tasks are somewhat complex and time consuming. The task gets even more complicated when there are different start and end times and when the auctions employ different protocols. For this reasons, a variety of software support tools are provided either by the online auction hosts or by third parties that can be used to assist consumers when bidding in online auctions.

The software tools include automated bidding software, bid sniping software, and auction search engines. Automated bidding software or proxy bidders act on the bidder's behalf and place bids according to a strict set of rules and predefined parameters. Bid sniping software, on the other hand, is a practice of placing of bid a few minutes or seconds before an auction closes. These kinds of software, however, have some shortcomings. Firstly, they are only available for an auction with a particular protocol. Secondly, they can only remain in the same auction site and will not move to other auction sites. Lastly, they still need the intervention of the user, that is, the user still needs to make decision on the starting bid (initially) and the bid increments.

To address the shortcomings mentioned above, an autonomous agent was developed that can participate in multiple heterogeneous auctions. It is empowered with trading capabilities and it is able to make purchases autonomously (Anthony, 2003; Anthony & Jennings, 2003b). Two primary values that heavily influenced the bidding strategies of this agent are the *k* and *β*. These two values correspond to the polynomial function of the four bidding constraints, namely the remaining time left, the remaining auction left, the user's desire for bargain and the user's level of desperateness. Further details on the strategies will be discussed in Section 3. The *k* value ranges from 0 to 1 while the *β* value is from 0.005 to 1000. The possible combinations between these two values are endless and thus, the search space for the solution strategies is very large. Hence, genetic algorithms were used to find the nearly optimal bidding strategy for a given auction environment.

This work is an extension of the solution above, which has been successfully employed to evolve effective bidding strategies for particular classes of environment. This work is to improve the existing bidding strategy through the optimization techniques. Three different variations of genetic algorithm techniques are used to evolve the bidding strategies in order to search for the nearly optimal bidding solution. The three techniques are parameter tuning, deterministic dynamic adaptation, and self-adaptation. Each of this method will be detailed in Section 4, 5 and 6. The remainder of the chapter is organized as follow. Section 2 discusses related work. The bidding strategy framework is discussed in Section 3. The parameter tuning experiment is discussed in Section 4. Section 5 and 6 discussed the deterministic adaptive experiment and self-adaptive experiment. A comparison between all the schemes is discussed in Section 7. Finally, the conclusion is discussed in Section 8.
