MODULAR OF WEIGHTLESS NEURAL NETWORK ARCHITECTURE FOR MOBILE ROBOT

Transcription

1 MODULAR OF WEIGHTLESS NEURAL NETWORK ARCHITECTURE FOR MOBILE ROBOT Siti Nurmaini 1, Siti Zaiton Mohd Hashim 2, A. Zarkasih 3 1,3 Department of Computer Engineering, Faculty of Computer Science, University of Sriwijaya, Indonesia 2 Department of Software Engineering, Faculty of Computer Science and Information System,UTM, Johor Bahru siti_nurmaini@unsri.ac.id 1, sitizaiton@utm.my 2, zarkasih@unsri.ac.id 3 ABSTRACT Inaccurate sensors, world unpredictability and imperfect control often cause the failure of traditional navigational methods for real time mobile robots. In face of these problems adaptive systems are imperative. Neural networks have been extensively used in the resolution of these. However, the standard multi layer perceptron (MLP) type network has various drawbacks, one of which is training requires repeated presentation of training data, which often results in very long learning time. An alternative type of network, almost unique, is the Weightless Neural Network (WNNs) these are also called RAM based networks. This paper describes WNNs for detection obstacle and recognizes the environment using a simple microprocessor system. In this paper we use modular architecture to enhance generalization and look-up table to minimize the execution time. The output stored into the robot RAM memory and becomes the current controller that drives the robot. This functionality is demonstrated on a mobile robot using a simple, 8 bit microcontroller with 512 bytes of RAM. The modular WNNs approach is code efficient only 10 Kbytes of source code, works well, and the robot was able to successfully recognize the obstacle in real time. Keywords: Modular architecture, weightless neural network, environmental recognition, microprocessor system, mobile robot 1. INTRODUCTION The neural computing models described in this paper are based on artificial neurons which have binary inputs and outputs and no weight between nodes. Neuron functions are stored into look-up tables, which can be implemented using commercially available Random Access Memories (RAMs). Instead of adjusting weights learning on Weightless Neural Networks (WNNs) generally consists of changing the contents of look-up table entries, which results in highly flexible and fast learning algorithms. In that approach we can use supervised and unsupervised learning techniques have been used with WNNs [10-12]. This research concerns development of efficient WNNs for real time detection of obstacles in a mobile robot based on a simple microprocessor system. Some work done by others indicates this approach may he successful. Mitchell et al. [2] has shown that a simple Hopfield network can be used to control an insect robot s movement, which can move forwards, backwards or move around an obstacle, by using only a Z80-based microcontroller system. In an unpublished paper by Zhou et al. [3], a simple perceptron network was developed to control a fire-fighting robot. The neural network was used to detect obstacles, find a fire, and then the robot moved to extinguish the fire. Yao et al. [9] present a RAM base neural network for a mobile robot using a simple microprocessor system, the method allows the robot to detect and avoid obstacles in real time. Despite the chosen technique, before implementing a neural network it is necessary to perform some tests to determine the network architecture (number of neurons and connectivity) that can solve the intended problem with the best relation between cost and recognition level. A software tool was developed to help with this task and the employed methodology used to choose a specific RAM topology. The software implementation is more flexible and portable. However, software implementation of RAM-based ANN is limited by the amount of available memory since RAM neurons are assembled as binary vectors, addressed by the inputs and stored into system memory [7]. If the number of inputs is high, what is very common in complex problems, each neuron will need a great amount of memory. This fact can make intractable the application of the RAM model in some cases with less available memory like microcontrollers. Although, the software solutions in WNNs many times faster than MLP neural network[7], but the software implementation is not as fast as the circuit hardware implementation due to the serial execution of the logic operations of the neurons by the microprocessor that takes many clock cycles in 237

2 238 The 5 th International Conference on Information & Communication Technology and Systems comparison to the hardware approach, which takes just one. The advantage of this technique is the employment of the microprocessor ALU to execute the logic operation set of the RAM algorithm. Simple logic functions (NOT, AND, and OR) are faster to execute than complex floating-point operations, used by a greater number of neural models. Consequently, the network circuit can be directly executed into the microprocessor ALU, this fact decreases the execution time of the algorithm. So in this paper we proposed the combination of hardware and software in WNNs. We can use a logic minimization that reduces processing time and improves software performance. These networks are typically used in image-recognition applications because of their small size and computational requirements [4]-[5]. In general, even an intelligent control algorithm may not work well for a new environment in which it has not been trained. This is called the generalization problem. But the advantage of WNNs is their modularity. This characteristic simplifies the modification of the architecture. The number of neuron inputs can be modified by rearranging the connectivity to the sensors alone. In this paper we use the idea of modularity. The modular WNNs architecture to propose for enhance generalization. This work demonstrates the potential of neural networks in embedded applications where powerful computers may not be available. In the paragraphs that follow, the robot s architecture and sensor system and the weightless neural network structure will be presented. Results of experiments that measure the robot s collision-avoidance ability will also be given. 2. MODULAR WNNs 2.1. Concepts of Modularity Modularity can be viewed as a manifestation of the principle of divide and conquer, which allows us to solve complex problems, by dividing them into smaller sub problems, easier to conquer, combining their individual solutions to achieve the final solution. On of the greater improvements achieved by modular One of the greater improvements achieved by modular neural networks is better generalization. The convention in neural networks is to use an architecture as small as possible to obtain better generalization, because the ability of feed-forward neural networks to generalize is inversely proportional to the number of weights involved (Baum, Haussler, 1989). Some other advantages occur by the use of modular neural networks. First, dividing a task into sub tasks will avoid the problems of temporal and spatial crosstalk. Temporal crosstalk occurs when a single neural network is trained in two different moments in time to perform two different tasks. Usually the net forgets how to execute the first task. Spatial crosstalk occurs when a neural network is trained simultaneously for the two tasks, resulting in incoherent training patterns. Finally, modularity can enable a learning economy, since some modules can be reused or modified without altering the other modules. This economy can also become crucial when hardware implementations are an objective, since modular neural networks can facilitate hardware execution of neural networks (Auda, Kamel, 1999). There are several types of modular networks: cooperative, competitive, sequential and supervisory (Sharkey, 1999). In all of them there is an implicit division of tasks. The difference among them is the way the modules interact. While in cooperative modular networks all modules cooperate in the final decision, in competitive networks the modules compete to win the chance to provide the solution. In sequential modular neural networks the modules solutions are used sequentially to achieve the final solution, the computation of one module depending of the preceding one. Supervisory networks have the task of supervising one or more modules (Sharkey, 1999) RAM base WNNs The basic architecture is as follows, the input vector is divided into parts; each part is connected to the address inputs of a 1 Bit-RAM unit. The outputs of all the RAMs within one discriminator are summed up. Figure 1. a Single RAM Unit The 1-Bit-RAM unit, depicted in Figure 1 is a device which can store one bit of information for each input address. A control input is available to switch the mode of the RAM between Write and Read for learning and recall. Initially all memory units are set to 0. During the learn (Write) mode

3 C36 - Modular Of Weightless Neural Network Architecture For Mobile Robot - Siti Nurmaini 239 the memory is set to 1 for each supplied address; in the recall (Read) mode the output is returned for each supplied address, either 1 (if the pattern was learned) or 0 (if the pattern was not learned). The discriminator is the device which performs the generalization. It consists of several RAMs and one node which sum the outputs of the RAMs in recall mode. The discriminator is connected to the whole input vector; each RAM within the discriminator is connected to a part of this vector, so that each input bit is connected to exactly one RAM, see Figure 2. The connections are preferably chosen by random. Figure 2. RAM network Neural network base multi-ram discriminator as shown in Figure 3; the system is built by grouping together a set of discriminators, each one being responsible by recognizing a different class of patterns. In the diagram below weightless neurons are connected to form such a discriminator. The neuron inputs are connected in a random sequence to the feature vector, each neuron is a binary pattern recognition device. Figure 3. Discriminator Each discriminator consists of M RAM-like neurons (weightless neurons or n-tuples) with n address lines, 2n storage locations (sites) and 1-bit word length. Each RAM randomly samples n bit of the input pattern, as shown in Figure 2. Each pattern must be sampled by at least one RAM. For an input vector of size K, the number of necessary neurons J of connectivity N that should be used to cover all inputs of the input vector should satisfy: J N > K. This neuron group is called a WISARD discriminator and its response is produced by connecting an adder that sums the neuron outputs, counting the number of active neurons (neurons outputting 1 ) in the group [6]. This response vector can be regarded as a feature vector that measures the similarity of an input pattern to all classes. In the WNN (n-tuple networks), a Winner- Takes-All can be attached to the adder outputs to choose the discriminator containing the greater number of active neurons, pointing to the winning class, each pattern will produce a feature vector that describes its similarity to all classes. The set of feature vectors generated by the WNN for a given training set will be used as input data to the fuzzy rule-based system. 3. THE ROBOT SYSTEM 3.1. The Architecture The system of the robot in Figure 4 is a modular system and containing several microcontrollers implies the implementation of a robust communication mechanism between modules. For this platform, the architecture can conceptually be seen as the central module, the motor driver module and the sensor module are connected to each other via 8 bit data bus. Each microcontroller can be equipped with a central control module as the motion unit controller. We use microcontroller AT89x55 for central controller, attached with ROM 20 Kbytes, RAM 256 bytes and clock 24.3 MHz, operated 0.5 micro-second for each process. A general purpose high performance controller for decision making is connected to the parallel bus. There are eight ultrasonic range finder sensors located at the front, left and right side of the robot. For sensor module, eight PIC16F84 chips are used for process signal detection from ultrasonic sensors, with ROM 1 Kbytes. Attached to the mobile robot, each separated at 30 o along the circumference. Furthermore, for real time control applications, PIC18F2550 is used exclusively to generate the Pulse Width Modulation (PWM) signals and to run the two dc motors, attached to each motor is an optical encoder which is used for distance and velocity calculation. Figure 4 illustrates the architecture that is available on the respective modules.

4 240 The 5 th International Conference on Information & Communication Technology and Systems Sensor 1 Sensor 2 Sensor 3 Sensor 4 Sensor 5 Sensor 6 Sensor 7 Sensor 8 SESOR MODULE PIC 16F84 MCU1 MCU2 MCU3 MCU4 MCU5 MCU6 MCU7 MCU8 Figure 4. Hardware architecture For this platform, the central control module for implements the algorithms of the navigation control. The sensor module is configured for controls the ultrasonic pulses, calculates the sensor readings, and selects which sensors are active according their position in the robot periphery. The motor drive module is configured the speed levels of the robot. The neural network classifies the environment using information from sensor module into geometric features such as U-shape, corridor and left or right corner and utilizes previous sensor data to analyze the situation for making navigation decision Sensor Position Parallel Bus Data 8 bit AT89X55 The robot behaviors tested by eight sensors, and divided into four behaviors namely obstacle avoidance, left and right wall following, and emergency condition. Figure 5, show sequence of measured range values to the obstacle and follow the wall. The sensors position in all array are fixed: one on back sides, and seven facing outwards at 30- degree intervals starting from 1-7. They report the distance between the robot and the closest obstacle. Figure 5. Position of the sensors on the robot The placement of the sensors permits the detection of an obstacle in different positions. That configuration was very efficient during preliminary tests of robot navigation. Seven sensors positioned CS CENTRAL MODULE AT89X55 Infra Red Module MOTOR PIC 18F255 DRIVER M in the front of the robot, considering that the robot will be moving when it is maneuvering, there is no situation where a robot can approach an obstacle without seeing it. One sensor is placed in the back to allow the robot to detect emergency condition and the robot must be move backward or stop. 4. MODULAR WNNs CLASSIFIER 4.1. Environment Classification We considered the common target that there exist in real environment of mobile robot applications such as plane, edge, corner with angle 90 degree, acute corner with angle 60 degree. Length of plane is about 45 cm and other objects are of similar size. These objects at four distances: 10, 20, 30, 40 cm. Also angle between the head of mobile robot and these objects is assumed to be -30, -20, -10, 0, 10, 20, 30 degree. Figure 6, show the environmental classification at these nine classes environment. Figure 6. Environmental classification 4.2. Structure of WNNs The WNNs is designed to identify the current environment by recognizing typical patterns. To implement this network, the 8 bit data from eight ultrasonic sensors is used to determine the direction of the obstacle, in each direction an obstacle may or may not appear. The combination of them appearance in the seven directions makes up different input pattern. In this experiment, a twolayer RAM base neural network was used, there are sensor layer and output layer, the structure is shown in Figure 7. The neural network used eight neurons for identifies where obstacle may lie in seven direction through ultrasonic sensor and seven output commands means that the neural network needs eight bits to encode them. Its configuration has seven groups of eight neurons (m=8 and n=7), so the network has 56 neurons. The winner-takes-all

5 C36 - Modular Of Weightless Neural Network Architecture For Mobile Robot - Siti Nurmaini 241 decision chooses the command that has more active neurons. The output layers have seven neurons, there are seven position of obstacle in the robot environment i.e.: obstacle in the front, one in front right another in front left of the robot or the opposite obstacle in right side another in right front of the robot obstacle in left side another in the left front of the robot obstacle on the left side another in the right side of the robot obstacle on the right side another in the left side of the robot obstacle on the multiple side of the robot obstacle in the front of the robot very near or emergency condition input vector Xn RAM RAM RAM RAM RAM RAM r1 r2 rk Decision - unit R conf value distinguishing distance an obstacle has to be obtained first. In the evaluation phase, the RAMbase network work by generating all the possible input combination. The number of possible combination of the eight sensors is 2 8 = 256 combination. Then, each output for all input possibilities was written to a lookup table, representing the neuron combination. The 8-bit signal from the sensors is directly connected to the address lines of the memory. The calculation of this value is based on the distance from an obstacle to the robot, as shown in Figure 8. Based on this calculation, the threshold values for the robot are (30 cm) indicates the obstacle is far, (20 cm) is medium, and (10 cm) is near and (75 cm) no obstacle is detected. 5. EXPERIMENTAL RESULT Experiment is conducted to demonstrate the ability of a mobile robot to recognize to various unknown environment, for geometric feature such as U-shape, corridor and left or right corner, particularly the ability of a robot to escape from the trapping situation. The base address 30H in hexadecimal is the address in memory of the first byte of the 256 locations of the memory. Table 1. Environmental Recognition 0 RAM RAM RAM Figure 7. Structure of WNNs (a) U-Shape (b) Left Wall (c) Right Wall (d) Open Space (e) Wide Corridor (f) Right Corner (g) Narrow Corridor (h) Left Corner (i) Isolated Obstacle Figure 8. Choosing the training value 4.3. Learning Strategy Here we use patterns with a single far, single medium, or single near obstacle to train the neural network. At a time the obstacle is placed in different directions and in difference distance. The neural network then is taught that the obstacle at left side, right side or forward. Using this technique, the Actual Place Convex (90 o ) Concave (270 o ) Font plane (180 o ) Left-corner (180 o ) Right-corner (180 o ) Corridor (0 o ) U-shape (180 o ) Distance (cm) Refe- Rence (hex) Result (hex) 20 12h 12h 30 0bh 0bh 40 05h 0dh 10 06h 00h 20 0eh 00h 30 00h 00h 10 03h 07h 10 03h 07h 10 03h 07h 10 03h 07h 10 03h 03h

6 242 The 5 th International Conference on Information & Communication Technology and Systems The algorithm WNNs classifier has been implemented in different environments. The system was capable of learning in simple problem. Also the system could generalize between different environments in complex problem as long as the encountered obstacles were similar. The result is based on the environment classification can show in Table 1 and Table 2. In this paper RAM network decision based on competitive learning algorithm and using 10 experiment data has achieved 95 % recognition for obstacle and 94 % classification. However the poorest result was if the robot closes the object, where the scanning sensory sector of the robot was quite high and some noise has still interfered in echo signal. Target angle Table 2. Obstacle Recognition Obstacle Obstacle Plane Wall Cylinder angle distance (cm) 10 0eh= 0eh= 0eh=14d 14d 14d 20 12h= 12h= 12h=18d 18d 18d 30 17h= 17h= 17h=23d 23d 23d 40 1dh= 1dh= 1dh=29d 29d 29d 10 0ch= 0bh= 0ch=12d 12d 11d 20 16h= 17h= 16h=22d 20 22d 23d 30 20h= 22h= 20h=32d 32d 34d 40 2dh= 2eh= 2dh=45d 45d 46d eh= d 40 2ah= 42d The Modular structure of WNNs proved to be efficient; thus showing generalization after learning. From the experiment controller classifies immediate environment by recognizing the pattern that were encountered during training 6. CONCLUSION Strategy has been developed which allows mobile robot to learn and move around an environment avoiding obstacles and this has been implemented in hardware. The operation of the complete controller that would take hundreds of lines of code to be implemented can be executed with one single instruction that reads the command byte in the memory. It can make the controller work hundreds of times faster. The controller is stimulated with every possible input and a corresponding output table is written. This output table is then stored into the robot RAM memory and becomes the current controller that drives the robot in the current environment. The network was implemented with very modes microcontroller system 20 Kbytes ROM, small amount of data memory about 512 bytes and 10 Kbytes source code program using the microcontroller assembler and C code and the robot was able to detect and avoid obstacles in real time. REFERENCES [1] Simoes, E.D.V., An Embedded Evolutionary Controller to Navigate a Population of Autonomous Robots, Frontiers in Evolutionary Robotics, I-Tech Education and Publishing, Vienna, Austria Book edited, ISBN , pp. 596, April [2] R. J. Mitchell, D. A. Keating, and C. Kambhampati, Neural network controller for mobile robot insect, Internal report, Department of Cybemetics, University of Reading, April 1994 [3] Y. Zhou, D. Wilkins, R. P. Cook, Neural network for a fire-fighting robot, University of Mississippi [4] I. Aleksander, W. V. Thomas, P. A. Bowden, Wisard: A radical step forward in image recognition. Sensor Review, pp , July 1984 [5] J, Austin, A Review of RAM-Based Neural Networks. Proceedings of the Fourth International Conference on - MICRONEURO94, pp ,1994 [6] L. Teresa et al, Weightless Neural Models: A Review of Current and Past Works, Neural Computer Surveys v. 2,pp41-61, 1999 [7] Simoes, E. D. V., Uebel, L. F., and Barone, D. A. C., Hardware Implementation of RAM Neural Networks. In Pattern Recognition Letters, n. 17, pp , 1996 [8] Botelho, S. C., Simoes, E. D. V., Uebel, L. F., and Barone, D. A. C., High Speed Neural Control for Robot Navigation. Proceedings of the IEEE International Conference on Systems, Man and Cybernetics, Beijing, China, pp , [9] Q. Yao, D. Beetner, D.C. Wunsch, and B. Osterloh., A RAM-Based Neural Network for Collision Avoidance in a Mobile Robot, IEEE, 2003 [10] J.Austin and T. J. Stonham, An associative memory for use in image recognition and occlusion analysis, Image and Vision Computing, v. 5, pp , 1987

7 C36 - Modular Of Weightless Neural Network Architecture For Mobile Robot - Siti Nurmaini 243 [11] E.Filho, M.C.Fairhurst and D.L.Bisset Adaptive pattern recognition using goalseeking neurons, Pattern Recognition Letters, v.12, , march 1991 [12] S. Ramanan, R.S. Petersen, T.G.Clarkson and J.G.Taylor pram nets for detection of small targets in sequence of infra-red images, Neural Networks, pages , 1995

8 244 The 5 th International Conference on Information & Communication Technology and Systems [This page intentionally left blank]