Volume 8 Issue 3 - April 3, 2009
A symbol-based intelligent control system with self-exploration process
Liang-Hsuan Chen1,* and Cheng-Hsiung Chiang2

1Department of Industrial and Information Management, National Cheng Kung University
2Department of Computer Science, Hsuan Chuang University

Engineering Applications of Artificial Intelligence, Vol. 21, No. 2, pp. 201-214 (2008)

Font Normal   Font Enlarge
The concept of human “intelligence” has been widely discussed in social science. Moreover, it has also been investigated in computer science, engineering, medicine, and management. In the fields of computer science and engineering, an intelligent system has the ability to sense the environment, to make decisions, and to generate actions. Thus, an intelligent control system (ICS) has three basic functions: perception, decision-making, and action. One can find many applications of the ICS in our life, such as a fuzzy-based air conditioner, a pilotless plane, an intelligent robot, etc. Generally, a control system can be divided into two types, i.e. indirect control and direct control. The indirect control system requires models to design the controller systems; however, the direct control system, such as the artificial intelligence technique (AIT) models, does not need any models for the controlled system. In this study, an AIT model is considered to implement an ICS. We present a new model of an ICS that can adapt to various environments through adjusting its control behaviors, namely SyICS (a Symbol-based Intelligent Control System).

The SyICS is comprised of a symbolic controller, a percepter and a self-adaptor, as shown in Fig. 1. In the figure, the symbolic controller produces the control inputs in the plant (controlled system), and then the percepter, which is analogous to the human sensorium, perceives the performance of the system. There are a number of symbolic rules, such as IF-THEN rules, in the symbolic controller. The percepter has two possible outputs: one connects to the symbolic controller and the other connects to the self-adaptor. If the system cannot adapt to the environment well or its efficiency is poor, the self-adaptor will be activated. Once the self-adaptor is activated, the self-exploration process implemented by the hybrid genetic algorithm (HGA) will start to explore the new control rules. After the new rules are finding, they will be transformed to the symbolic rules through the symbolic rules generator. Following, percepter, self-exploration process, and hybrid genetic algorithm are briefly introduced.

Percepter: The percepter is regarded as a sensing mechanism in order to detect the control efficiency. In this study, we present a perceptive engine, namely the percepter, consisting of two measures, adaptability and efficiency measures. The adaptability measure: pa indicates whether the control system can adapt to the environment or not, and is defined as
The efficiency measure: pe indicates the control efficiency of system, and is to be normalized to [0, 1]. The overall measure of the percepter is defined as
whereηindicates the adjustable constant. If the output of percetper <ε, the self-adaptor will be activated; otherwise, the controller continues its controlling task.

Self-Exploration Process: An illustration of the self-exploration process is shown in Fig. 2. In the figure, the system first produces the six actions. However, the control system fails to produce the adaptable action at the 6th time step. Thus, the system goes backward to check its actions, and then returns two time steps to action Ad. Afterwards, the system decides to find seven new actions in order to adapt to the environment. Finally, the system finds out the seven new actions.

Hybrid Genetic Algorithms: In this study, we present a hybrid genetic algorithm (HGA) with five features: 1) discrete encoding, 2) variable length chromosomes, 3) weighted-sum based multi-objective fitness function, 4) multiple genetic operators, and 5) elitist population update strategy. In this study, five genetic operators are employed, namely selection, crossover, mutation, insertion and deletion operators.
Figure 1 Architecture of SyICS.
Figure 2 An illustration of the self-exploration process.

An Application of Robotic Path Planning

A simulated robot must navigate from a start position to a target destination in the designed environment, as shown in Fig. 3. The environment is a square plane with size 100×100 cm2. Suppose that the robot can recognize both its position and the target’s position, while the robot can only recognize an obstacle when a collision occurs.
Figure 3 An illustration for robotic path planning (a) Sketches of robot and target zone (b) Moving space of robot.

Simulation Results

In order to demonstrate the proposed approach, we compare the proposed SyICS model with the similar SEICS [1] and mSEICS [2] models. In the simulated environment, there are twenty-five obstacles, and the start and target points are (5, 15) and (50, 85), respectively, as shown in Fig. 4. The paired slanted dotted-lines in Fig. 4(b) show the tolerable area of the robot for the mSEICS method. There are two types of paths for each method in Fig. 4. One is the original path labeled by the circles (‘o’), and the other is the adaptive path labeled by the star symbols (‘*’). As shown in Fig. 4, all three methods have at least one collision. Fig. 4(a) shows that the SyICS approach has two collisions, and the self-adaptor is activated twice to search the new symbolic rules. As shown in Fig. 4, after updating the rules, the adaptive path shows that the robot first makes a detour toward the left side to avoid bumping into the obstacle, and then moves a little to the right to avoid the second obstacle. The adaptive path of mSEICS makes a larger detour to avoid these two obstacles, as shown in Fig. 4(b). The original and adaptive paths of SEICS model shown in Fig. 4(c) demonstrate a different strategy for passing the obstacle. The efficiency measures of the original and adaptive paths of the SyICS model are 85.07 cm and 77.14 cm, respectively. However, for mSEICS model and SEICS model, they have (255.74 cm, 35.76 cm) and (521.91 cm, 603.85 cm), respectively. Compared with the mSEICS and SEICS models, the SyICS model has the best efficiency. Moreover, the moving steps from the start point to the target point for the adaptive paths of SyICS, mSEICS and SEICS are 29, 43 and 43, respectively. Obviously, the SyICS model requires the minimum number of moving steps in this case.
Figure 4 Simulation results with twenty-five obstacles. (a) SyICS method (b) mSEICS


The above simulation results showed that 1) the robotic path of the SyICS model is the most efficient based on the path’s efficiency; 2) the number of instances adaptive learning required by SyICS to find a successful path without collision is the minimum; 3) the paths produced by the SyICS model are closest to the shortest path when compared with the mSEICS and SEICS models. Therefore, the comparison demonstrates that the performance of SyICS model is best among the three methods.


[1]L.-H. Chen and C.-H. Chiang, “New approach to intelligent control systems with self-exploring process,” IEEE Transactions on Systems, Man, and Cybernetics, Part B: Cybernetics, vol. 33, no. 1, pp. 5666, 2003.
[2]L.-H. Chen and C.-H. Chiang, “An intelligent control system with a multi-objective self-exploration process,” Fuzzy Sets and Systems, vol. 143, no. 2, pp. 275294, 2004.
< previousnext >
Copyright National Cheng Kung University