Boontee Kruatrachue

HMM topology selection for on-line Thai handwriting recognition

Researchers have extensively applied hidden Markov models (HMM) to handwriting recognition in English, Chinese, and other languages. Most researchers have used left-right topology for handwriting and speech recognition. This research studied the effect of HMM topology on isolated online Thai handwriting recognition. The left-right, fully connected and proposed topologies (left-right-left) were compared. The number of states of a character HMM for each topology was varied from 15 to 35 nodes and the one with the best training observations probability was selected. The feature used was chain code-like with modifications to represent original quadrant position. The recognition results showed that the proposed topology increases the recognition rate compared to the most widely used left-right topology.

Online Unconstrained Handwritten Thai Character Recognition Using Multiple Representations

This paper proposes an unconstrained handwritten Thai character recognition method using multiple representations. The proposed method can recognize different handwritten Thai characters having small curve segments with clockwise and counter clockwise directions, although the conventional method such as the elastic matching method was difficult to recognize them. As the experimental results, the recognition rate was 97.50% and the recognition time (average) using the PC with 1.4 GHz Pentium 4 was 0.01 second.

Tree-structured model selection and simulated-data adaptation for environmental and speaker robust speech recognition

This paper proposes the use of tree-structured model selection and simulated-data in maximum likelihood linear regression (MLLR) adaptation for environment and speaker robust speech recognition. The objective of this work is to solve major problems in robust speech recognition system, namely unknown speaker and unknown environmental noise. The proposed solution is composed of two components. The first one is based on a tree-structured model for selecting a speaker-dependent model that best matches to the input speech. The second component uses simulated-data to adapt the selected acoustic model to fit with the unknown noise. The proposed technique can thus alleviate both problems simultaneously. Experimental results show that the proposed system achieves a higher recognition rate than the system using only the input speech in adaptation and the system using a multi-conditioned acoustic model.

Adaptive parameter forecasting for FOREX automatic trading system using fuzzy time series

Efficiency of FOREX automatic trading system via technical indicator depends on suitability of its parameters. Using parameter forecasting with fuzzy time series is a simple and efficient solution. But its effectiveness depends on the similarities of the time series pattern to learn and to forecast which represent in the form of fuzzy relationship. This paper presents a method for time-varying fuzzy relationship. The forgetting factor allows the system to perform better in all time periods. Effects of the length of time series on forecasting the performance of the system, as well as those of profits derived from various methods on predicting the trading decisions in real markets are studied.

State Machine Induction with Positive and Negative Training for Thai Character Recognition

One problem of generating a model to recognize any string is how to generate one that is generalized enough to accept strings with similar patterns and, at the same time, is specific enough to reject the non-target strings. This research focus on generating a model in the form of a state machine to recognize strings derived from the direction information of characters images. The state machine induction process has two steps. The first step is to generate the machine from the strings of each target character (positive training), and the second step is to adjust the machine to reject any other string (negative training). This state machine induction method that automatically learns from strings can be applied with other string patterns recognition apart from characters.

Hybrid Genetic Algorithm with Baum-Welch Algorithm by using Diversity Population Technique

Baum-Welch Algorithm (BWA) have used in recognition systems, many researchers have improved BWA performances by using hybrid genetic algorithm (HGA). This paper presents a new HGA technique by using diversity population structure. We surveyed HGA techniques and divided into four types. There were separate processes, population types, fitness determiners, and diversity population structure. A technique of diversity population structure protected applying BWA to similar population. Different population structures make available GA to find optimum point quickly. This paper compared all of HGA techniques, which there trained on hidden Markov models (HMM), in an application Thai off-line handwritten recognition, we used database from NECTEC. An experiment of HGA, HMM probability of diversity population techniques get better than techniques of population types 52.65% improvement and there better than techniques of separately process 37.91% improvement. Moreover, HGA experimented five times repeatedly, standard derivation value of diversity population techniques showed closely results

Thai OCR error correction using genetic algorithm

This paper presents an efficient method for Thai OCR error correction based on genetic algorithm (GA). The correction process starts with word graph construction from spell checking with dictionary, then a graph is searched for a corrected sentence with the highest perplexity (using language model, bi-gram and tri-gram) and word probability from OCR. For a long sentence, a search space is huge and can be resolved using GA. A list of nodes is used for chromosome encoding to represent all possible paths in a graph instead of standard binary string. The performance of the suggested technique is evaluated and compared to the full search for tested sentences of different size constructed from 10 nodes to 200 nodes word graphs.

A modified multi-swarm optimization with interchange GBEST and particle redistribution

The Particle Swarm Optimization (PSO) is an optimization algorithm using multiples particle to search solution space for an optimize solution. Each particle of PSO moves toward the best solution within its group. For this behavior, PSO often traps in local optimum. Many researchers proposed splitting a swarm into multiple swarms so that they may move to different local optimum. Besides, the mutation operation technique, the natural selection technique and the crossover operation technique are added to normal PSO process. These proposed techniques are called Selective Crossover base on Fitness in Multi-Swarm Optimization (SFMPSO) and Fast Multi-swarm Optimization (FMPSO). However, both techniques used too many evaluation calls dues to crossover and the mutation operation. This paper proposes setting the best position (GBEST) of a trapped swarm to GBEST of the other swarm. Then, the swarms particle is redistributed in solution space before restart the trapped swarm. This proposed technique is evaluated on a set of twenty-six benchmark test functions. The experimental results show that the results are better than those of PSO, FMPSO and SFMPSO.

A Modified Particle Swarm Optimization with Dynamic Particles Re-initialization Period

The particle swarm optimization (PSO) is an algorithm that attempts to search for better solution in the solution space by attracting particles to converge toward a particle with the best fitness. PSO is typically troubled with the problems of trapping in local optimum and premature convergence. In order to overcome both problems, we propose an improved PSO algorithm that can re-initialize particles dynamically when swarm traps in local optimum. Moreover, the particle re-initialization period can be adjusted to solve the problem appropriately. The proposed technique is tested on benchmark functions and gives more satisfied search results in comparison with PSOs for the benchmark functions.

Prototype selection using Reinforcement Learning and Minimal Consistent Subset Identification guide

This paper try to apply Reinforcement Learning (RL) to a task with large number of states. This usually is a difficult task since RL has less chance to visit all state or has enough number of visit to learn average reward accurately. Moreover, RL may not be able to learn or obtain any optimal solution as RL learn by averaging rewards from each action performing in each state. In order to alleviate this RL learning problem, any solution to a task such as, non-optimal algorithm or heuristics can collaborate with RL by using their knowledge to prune the non-optimal action in each state. This reduces search space of RL and helps it learn faster. A Minimal consistent subset problem is used as an example to demonstrate how RL can learn faster with the help of other heuristics.