Research Purpose

In this dissertation, improved safe reinforcement learning (ISRL) based self adaptive evolutionary algorithms (SAEAs) for neuro-fuzzy controller is proposed for improving not only the reinforcement signal designed but also evolutionary algorithms mentioned in Section 1.1. There are two parts in the proposed ISRL-SAEAs.

In the first part, self adaptive evolutionary algorithms (SAEAs) are proposed to solve the following problems: 1) all the fuzzy rules are encoded into one chromosome; 2) the number of fuzzy rules has to be assigned in advance; and 3) the population cannot evaluate each fuzzy rule locally. In this dissertation, the proposed self adaptive evolutionary algorithms (SAEAs) consist of three different evolution methods to provide different ways to solve the above problems.

the numbers of fuzzy rules should be decided automa the numbers of fuzzy rules should be decided automa

First of all, the hybrid evolutionary algorithm (HEA) with a TSK-type neuro-fuzzy controller is proposed, the proposed HEA determines the number of fuzzy rules automatically and processes the variable-length chromosomes. The length of each individual denotes the total number of genes in that individual. The initial length of each individual may be different from each other, depending on the total number of rules encoded in it. Individuals with an equal number of rules constitute the same group. Thus, initially there are several groups in a population. For keeping the best group in every generation, the elite-based reproduction strategy (ERS) is proposed. In the ERS, the best group can be reproduced many times for each generation. The advantages of the proposed HEA are summarized as follows: 1) it determines the number of fuzzy rules and tunes the free parameters of the neuro-fuzzy controller in a highly autonomous way. Thus, users need not give it any a priori knowledge or even any initial information on these. 2) It is applicable to chromosomes of different lengths. 3) It does not require precise training data for setting the parameters of the neuro-fuzzy controller.

Although the proposed HEA can determine the number of fuzzy rules automatically, all the fuzzy rules are encoded into one chromosome. Therefore, partial solution cannot be evaluated independently in the population. The partial solutions can be characterized as specializations. The specialization property ensures diversity and prevents a population from converging to suboptimal solutions. A single partial solution cannot “take over” a population since it must correspond with other specializations. For solving this problem, the secondary algorithm of the SAEAs is proposed. In the secondary algorithm of the SAEAs, a self adaptive group cooperation based symbiotic evolution (SAGC-SE) is proposed not only for solving the problem that all the fuzzy rules are encoded into one chromosome but also for letting the population evaluate each fuzzy rule locally. Therefore, in the proposed SAGC-SE, each chromosome represents only one fuzzy rule and an n-rules TSK-type neuro-fuzzy controller is constructed by selecting and combining n chromosomes from several groups. The initial information on these. 2) It is applicable to chromosomes of different lengths. 3)

not require precise training data for setting the parameters of the neuro-fuzzy controller.

Although the proposed HEA can determine the number the fuzzy rules are encoded into one chromosome. Th

not require precise training data for setting the parameters of the neuro-fuzzy controller.

not require precise training data for setting the parameters of the neuro-fuzzy controller.

initial information on these. 2) It is applicable to chromosomes of different lengths. 3) not require precise training data for setting the parameters of the neuro-fuzzy controller.

Although the proposed HEA can determine the number

arameters of the neuro-fuzzy controller.

Although the proposed HEA can determine the number

arameters of the neuro-fuzzy controller.

Although the proposed HEA can determine the number not require precise training data for setting the p

Although the proposed HEA can determine the number Although the proposed HEA can determine the number not require precise training data for setting the p

not require precise training data for setting the p

Although the proposed HEA can determine the number Although the proposed HEA can determine the number not require precise training data for setting the p

not require precise training data for setting the p not require precise training data for setting the p not require precise training data for setting the p

Although the proposed HEA can determine the number Although the proposed HEA can determine the number not require precise training data for setting the p

Although the proposed HEA can determine the number

prevents a population from converging to suboptimal solutions. In SAGC-SE, there are several groups in the population. Each group formed by a set of chromosomes represents a fuzzy rule. The proposed SAGC-SE consists of structure learning and parameter learning. In structure learning, as well as HEA, the SAGC-SE determines the number of fuzzy rules automatically and processes the variable length of a combination of chromosomes. In parameter learning, to let the well-performing groups of individuals for cooperating to generate better generation, an elite-based compensatory of crossover strategy (ECCS) is proposed. In the ECCS, each group will cooperate to perform the crossover steps. Therefore, the better chromosomes of each group will be selected to perform crossover in the next generation.

The advantages of the proposed SAGC-SE are summarized as follows: 1) the proposed SAGC-SE determines the number of fuzzy rules automatically. 2) The SAGC-SE uses group-based population to evaluate the fuzzy rule locally. 3) The SAGC-SE uses the ECCS to allow the better solutions from different groups to cooperate for generating better solutions in the next generation.

The SAGC-SE can solve the problem of the HEA that all the fuzzy rules are encoded into one chromosome. Moreover the SAGC-SE evaluates each fuzzy rule locally for improving the local consideration of the population. However, in the SAGC-SE, how to select groups for constructing the complete solution is a major problem. Therefore, for determining the number of fuzzy rules automatically, the SAGC-SE selects different number of groups to construct complete solution. In this way, the SAGC-SE selects groups randomly. It’s obvious that the performance of the SAGC-SE dependents on the method of selecting groups. For solving this problem, in the third algorithm of the SAEAs, a self adaptive groups based symbiotic evolution using FP-growth algorithm (SAG-SEFA) is proposed.

As well as SAGC-SE, the SAG-SEFA consists of structure learning and parameter SAGC-SE determines the number of fuzzy rules automa

group-based population to evaluate the fuzzy rule locally. 3) The SAGC-SE uses the ECCS to allow the better solutions from different groups to cooperate for generating better solutions in group-based population to evaluate the fuzzy rule l

group-based population to evaluate the fuzzy rule l

SAGC-SE determines the number of fuzzy rules automa

group-based population to evaluate the fuzzy rule locally. 3) The SAGC-SE uses the ECCS to allow the better solutions from different groups to cooperate for generating better solutions in group-based population to evaluate the fuzzy rule l

allow the better solutions from different groups to group-based population to evaluate the fuzzy rule l allow the better solutions from different groups to group-based population to evaluate the fuzzy rule l allow the better solutions from different groups to allow the better solutions from different groups to group-based population to evaluate the fuzzy rule l group-based population to evaluate the fuzzy rule l allow the better solutions from different groups to allow the better solutions from different groups to group-based population to evaluate the fuzzy rule l group-based population to evaluate the fuzzy rule l group-based population to evaluate the fuzzy rule l group-based population to evaluate the fuzzy rule l allow the better solutions from different groups to allow the better solutions from different groups to group-based population to evaluate the fuzzy rule l allow the better solutions from different groups to

the number of fuzzy rules automatically and processes the variable combination of chromosomes. In parameter learning, although the proposed SAG-SEFA can determine the suitable number of rules, there still has a problem in which how to select the suitable groups from many groups (named candidate groups in this paper) in SAG-SEFA to construct TSK-type neuro-fuzzy controllers with different numbers of rules. Moreover, in consideration of making the well-performing groups of individuals cooperate for generating better generation, there is also a problem in which how to select suitable groups used to select individuals for cooperating to generate better generation. Regarding this, the goals of parameter learning in SAG-SEFA are used to determine which groups of chromosomes should be selected to construct TSK-type neuro-fuzzy networks with different numbers of rules and which groups should be selected for cooperating to generate better generation.

Recently, data mining has become a popular research topic ([45]-[48]). Data mining is a method of mining information from a database. The database called “transactions”. Data mining can be regarded as a new way of performing data analysis. One goal of data mining is to find association rules among sets of items that occur frequently in transactions. To achieve this goal, several methods have been proposed ([49]-[54]). In [49], the authors proposed a mining method which ascertains large sets of items to find the association rules in transactions.

Hang et al. ([50]) proposed frequent pattern growth (FP-growth) to mine frequent patterns without candidate generations. In Hang’s work, items that occur more frequently will have better chances of sharing information than items that occur less frequently. In [51], an algorithm of data mining for transaction data with quantitative values was proposed. In [51], each quantitative item was translated to a fuzzy set and the authors used these fuzzy sets to find fuzzy rules.Wu et al. ([52]) proposed a data mining method based on GA algorithm that efficiently improves the traditional GA by using analysis and confidence parameters. In [53], authors proposed a hybrid model using rough sets and genetic algorithms for fast and efficient

Recently, data mining has become a popular research method of mining information from a database. The d mining can be regarded as a new way of performing d to find association rules among sets of items that

method of mining information from a database. The d method of mining information from a database. The d Recently, data mining has become a popular research method of mining information from a database. The d mining can be regarded as a new way of performing d method of mining information from a database. The d mining can be regarded as a new way of performing d method of mining information from a database. The d mining can be regarded as a new way of performing d method of mining information from a database. The d mining can be regarded as a new way of performing d mining can be regarded as a new way of performing d method of mining information from a database. The d method of mining information from a database. The d mining can be regarded as a new way of performing d mining can be regarded as a new way of performing d method of mining information from a database. The d method of mining information from a database. The d method of mining information from a database. The d method of mining information from a database. The d mining can be regarded as a new way of performing d mining can be regarded as a new way of performing d method of mining information from a database. The d mining can be regarded as a new way of performing d

As shown in [53], authors proposed select, aggregate and classification based data mining queries to implement a hybrid model. The performance of the proposed algorithm is analyzed for both execution time and classification accuracy and the results obtained are good. In [54], Dai and Zhang proposed an association rules mining in novel genetic algorithm. The genetic algorithm in [54] are using for discovering association rules. As shown in [54], the proposed algorithm avoids generating impossible candidates, and it is more efficient than traditional ones.

Since data mining can successfully find information from large sets of items, it is useful to achieve goals of parameter learning in SAG-SEFA. Therefore, the data-mining method called FP-growth algorithm is adopted since the FP-growth algorithm can find items that occur frequently in transactions without candidate generations. After the parameter learning with FP-growth algorithm is performed, the population can search for a better solution from the combination of individuals that perform well and explore other combinations of individuals. Moreover, suitable groups will cooperate to perform the crossover steps.

Therefore, the better chromosomes of suitable group will be selected to perform crossover in the next generation.

When compared with SAGC-SE, the proposed SAG-SEFA not only selects the suitable groups form candidate groups to perform selection steps but also allows the better solutions from different groups to cooperate for generating better solutions in the next generation.

The second part of the proposed ISRL-SAEAs is an improved safe reinforcement learning (ISRL). In the ISRL, the feedback takes the form of an accumulator. The accumulator determines by two different strategies (judgment and evaluation). The judgment strategy determines the reinforcement signal when the plant fails entering a predefined goal set and the evaluation strategy applies under the condition that the plant enters the goal set. Moreover the safe reinforcement learning [32] is considered in ISRL. The key of the ISRL is using an with FP-growth algorithm is performed, the populati

the combination of individuals that perform well an individuals. Moreover, suitable groups will coopera Therefore, the better chromosomes of suitable group the combination of individuals that perform well an the combination of individuals that perform well an with FP-growth algorithm is performed, the populati the combination of individuals that perform well an individuals. Moreover, suitable groups will coopera the combination of individuals that perform well an individuals. Moreover, suitable groups will coopera the combination of individuals that perform well an individuals. Moreover, suitable groups will coopera the combination of individuals that perform well an individuals. Moreover, suitable groups will coopera individuals. Moreover, suitable groups will coopera the combination of individuals that perform well an the combination of individuals that perform well an individuals. Moreover, suitable groups will coopera individuals. Moreover, suitable groups will coopera the combination of individuals that perform well an the combination of individuals that perform well an the combination of individuals that perform well an the combination of individuals that perform well an individuals. Moreover, suitable groups will coopera individuals. Moreover, suitable groups will coopera the combination of individuals that perform well an individuals. Moreover, suitable groups will coopera

will be observed that the advantage of the proposed ISRL is that it can meet global optimization capability.