Christopher J. Gatti

Hierarchical Clustering for Large Data Sets

This chapter provides a tutorial overview of hierarchical clustering. Several data visualization methods based on hierarchical clustering are demonstrated and the scaling of hierarchical clustering in time and memory is discussed. A new method for speeding up hierarchical clustering with cluster seeding is introduced, and this method is compared with a traditional agglomerative hierarchical, average link clustering algorithm using several internal and external cluster validation indices. A benchmark study compares the cluster performance of both approaches using a wide variety of real-world and artificial benchmark data sets.

Forecasting exchange rates with ensemble neural networks and ensemble K-PLS: A case study for the US Dollar per Indian Rupee

The purpose of this paper is to evaluate and benchmark ensemble methods for time series prediction for daily currency exchange rates using ensemble feedforward neural networks and kernel partial least squares (K-PLS). Best-practice forecasting methods for the US Dollar (USD) per Indian Rupee (IR) are applied for training, validating, and testing the machine learning models. In order to perform the benchmarking evaluation study neural network forecasting methods are first compared on a benchmarked neural network time series prediction method for the Canadian Lynx time series. The K-PLS method is benchmarked in addition with support vector machines (SVM), a similar kernel-based method. Both one-step ahead and a roll-out methods for extended forecast horizons are applied for the currency exchange rates. The paper is novel in the sense that two new ensemble methods are introduced: weight seeding and multiple cross-validation averaging. The paper is also novel in the sense that several new validation indices are proposed that are especially applicable for time series: q2 and Q2 and the fraction of misses in the exchange rate return space, which is a more relevant metric for currency speculation. As a general conclusion it is found that the USD per IR is quite predictable, while other currencies such as the USD per Euro and the Australian Dollar (AUD) per Euro are not predictable.

Reinforcement Learning with Neural Networks: Tricks of the Trade

Reinforcement learning enables the learning of optimal behavior in tasks that require the selection of sequential actions. This method of learning is based on interactions between an agent and its environment. Through repeated interactions with the environment, and the receipt of rewards, the agent learns which actions are associated with the greatest cumulative reward.

The Mountain Car Problem

The mountain car problem is commonly used as a benchmark reinforcement learning problem to evaluate learning algorithms. The problem places a car in a valley, where the goal is to get the car to drive out of the valley (Fig. 5.1). The car’s engine is not powerful enough for it to drive out of the valley, and the car must instead build up momentum by successively driving up opposing sides of the valley. This chapter explores the mountain car problem using sequential CART and stochastic kriging to understand the parameter space.

Reinforcement learning and the effects of parameter settings in the game of Chung Toi

This work applied reinforcement learning and the temporal difference TD(λ) algorithm to train a neural network to play the game of Chung Toi, a challenging variant of Tic-Tac-Toe. The effects of changing parameters and settings of the TD(λ) and of the neural network were evaluated by observing the ability of the network to learn the game of Chung Toi and play against a ‘smart’ random player. This work applied techniques that have proven effective in training neural networks in general to the TD(λ) algorithm. The basic implementation of the TD(λ) method resulted in stable performance and achieved a maximal performance of winning 90.4% of evaluation games. When changing parameter settings, the best performance was achieved by using different learning rates between layers in the neural network (92.6% wins), and this was followed by using a relatively high probability of action exploitation (91.8% wins).

The Tandem Truck Backer-Upper Problem

The tandem truck backer-upper (TTBU) problem is an extension of the single trailer truck-backer upper problem where this problem uses a trailer truck with two trailers instead of one, which is considered to be nearly impossible for humans alone. A pure reinforcement learning approach has not be used to solve this problem, however, an assistive control scheme for this problem has been developed system that merely assists, but does not replace, the truck driver in backing up the tandem trailer. The work described herein is the first to use a simple reinforcement learning approach to begin to learn the tandem trailer-backer upper problem, and we explore the ability of the temporal difference algorithm to learn this problem in this chapter.

Design of Experiments

In this chapter, we review relevant concepts from the field of design of experiments, and this review assumes some basic knowledge of the field. We review both classical and contemporary design of experiments methods. Classical methods are well-established and have a long history of use in many applications; some of these include factorial designs, ANOVA (analysis of variance), and response surface modeling amongst others. The contemporary methods considered are those that are suited for design of experiments for computer simulations, which are based on some fundamental differences from classical experiments. These methods are primarily based on the experimental design and the creation of metamodels of response surfaces (i.e., surrogate models that could be use replacements for true computational models).

The Truck Backer-upper Problem

In this chapter, we focus on analyzing the truck backer-upper problem (TBU), a real world-like control problem. In this problem a tractor trailer truck must be backed into a specific location with a specific orientation by controlling the orientation of the wheels of the truck cab. We use sequential CART and stochastic kriging to understand how parameters of the neural network and learning algorithm affect convergence and performance in the TBU domain.

Erratum to Taguchi-fuzzy multi output optimization (MOO) in high speed CNC turning of AISI P-20 tool steel [Exp. Syst. Appl. 38(6) (2011) 6822-6828]

While utilizing an untested technique to extract additional information from existing data set, we found unusual results in 14 of the 107 papers so we carefully examined each paper. In this and thirteen other papers we found data and/or calculation errors. It is not clear if either some numbers were mis-transcribed while typesetting or if someone made a calculation error during the data collection stage or some other error occurred. It is our belief errors such as this are commonplace in scientific and industrial data, once an error has been identified it needs to be corrected, hence this note has been written. In addition to finding fourteen papers with evidence of errors, we found five data sets with errors that were a result of mis-entry of data by our students when preparing the analysis. Our process for data entry had a Masters candidate transcribe data into excel files, then a second graduate student checked and corrected all data files. In other words with 100% inspection by a masters student that was deemed reliable, 5 out of 107 data sets still contained errors. Below is a comparison of the data that appeared in the paper and our calculations (Tables 1 and 2).

Erratum to Multi-attribute decision making for green electrical discharge machining [Expert Syst. Appl. 38 (7) (2011) 8370-8374]

While utilizing an untested technique to extract additional information from existing data set, we found unusual results in 14 of the 107 papers so we carefully examined each paper. In this and thirteen other papers we found data and/or calculation errors. It is not clear if either some numbers were mis-transcribed while typesetting or if someone made a calculation error during the data collection stage or some other error occurred. It is our belief errors such as this are commonplace in scientific and industrial data, once an error has been identified it needs to be corrected, hence this note has been written. In addition to finding fourteen papers with evidence of errors, we found five data sets with errors that were a result of mis-entry of data by our students when preparing the analysis. Our process for data entry had a Masters candidate transcribe data into excel files, then a second graduate student checked and corrected all data files. In other words with 100% inspection by a masters student that was deemed reliable, 5 out of 107 data sets still contained errors. Below is a comparison of the data that appeared in the paper and our calculations (Table 1).