KR20010095569A - Genetic Programming Method and System using Process Decomposition - Google Patents

Abstract

PURPOSE: A genetic programming method using a process division method and a system therefor are provided, thereby accomplishing the adaptive system which can adapt to the changes of environment and duty. CONSTITUTION: The genetic programming method comprises the steps of: analyzing the given hardware resource; selecting a function nod and a vertical nod consisting of a tree; selecting a condition nod used as a root nod of sub-tree; evolving the tree using the function nod and vertical nod; dividing the evolved tree into an appropriate size; modifying the divided subtree to be suitable for the hardware; making the hardware circuit containing the divided subtree and the condition nod and expressing it to the hardware; calculating the final results by the sum of interim results; and performing the context switching by automatically dividing the process and converting it.

Description

Genetic Programming Method and System Using Process Decomposition

Genetic programming using a process partitioning and a system using the same according to the present invention is a method of converting a gene tree having a target function using gene programming and then splitting the transformed gene tree into a hardware representation as a context conversion (context The switching method is introduced.

In the context switching method, an evolutionary operation divides a large process into a set of small processes, converts the partitioning result into a hardware-implementable form, executes the divided subprocesses according to the calculation order, and performs intermediate calculation. The result is the sum of the results.

Genetic programming using process partitioning and the system using the same according to the present invention uses genetic programming as a dynamic reconstruction method of evolutionary hardware.

Evolutionary hardware refers to hardware that is designed to automatically adapt to changes in the environment or missions using runtime dynamic reconfiguration capabilities. Most previous studies have used genetic algorithms for control of evolutionary hardware.

Naito created a control circuit of an autonomous mobile robot on a simulation using a genetic algorithm (T. Naito, R. Odagiri, Y. Matsunaga, M. Tanifuji, and K. Murase, Genetic Evolution of a Logic Circuit). Which Controls an Autonomous Mobile Robot, First International Conference on Evolvable Systems 1996, pp. 210-219, 1996.).

Iba evolved hardware circuits by considering the configuration bit stream of the PLD as an entity of genetic algorithms (H. Iba, M. Iwata, and T. Higuchi, Machine Learning Approach to Gate-Level Evolvable). Hardware, First International Conference on Evolvable Systems 1996 , pp. 327-343, 1996.).

Miller evolved 3-bit binary multipliers using XILINX xc6216 FPGAs (JF Miller, and P. Thomson, Aspects of Digital Evolution: Geometry and Learning, Second International Conference on Evolvable Systems 1998 , pp. 25-35 , 1998).

Evolutionary hardware researchers prefer genetic algorithms for the following reasons: First, genetic algorithms can use the configuration bit stream of the underlying hardware as an entity. Second, genetic algorithms are simple and practical. Third, the hardware implementation of the genetic algorithm is simpler than the hardware implementation of the genetic programming.

However, there was no attempt to use genetic programming.

Koza used evolutionary hardware to evolve the sorting network. Koza has implemented parallel computational circuits for fitness evaluation on evolutionary hardware (JR Koza, FH Bennett III, JL Hutchings, SL Bade, MA Keane, and D. Andre, Evolving computer programs using rapidly reconfigurable field-). programmable gate arrays and genetic programming, Proceedings of the ACM Sixth International Symposium on Field Programmable Gate Arrays , pp. 209-219, 1998).

Bennett III used gene programming for the evolution of 60 decibel op amps (FH Bennett III, JR Koza, D. Andre, and MA Keane, Evolution of a 60 Decibel Op Amp Using Genetic Programming, First International Conference on Evolvable Systems 1996 , pp. 455-469, 1996).

Sakanashi has improved the accuracy of circuit representation using genetic programming in the hardware implementation of binary decision diagrams (H. Sakanashi, T. Higuchi, H. Iba, and Y. Kakazu, Evolution of Binary Diagrams for Digital Circuit). Design Using Genetic Programming, First International Conference on Evolvable Systems 1996 , pp. 470-481, 1996).

Compared to genetic algorithms, the advantage of genetic programming is that the functional nodes of the tree object can be used directly as functional blocks in hardware circuits. Thus, a search method based on genetic programming uses function blocks in randomly distributed binary strings. You can avoid unnecessary search processes that evolve. However, due to the difficulty of hardware implementation due to the large individual size of gene programming, a new method for hardware application is required. In addition, many researchers have tried to make autonomous robots based on various algorithms.

Nolfi (S. Nolfi, Using Emergent Modularity to Develop Control Systems for Mobile Robots, Adaptive Behavior vol. 5, pp.343-363, 1997) has produced excellent garbage collection robots based on modular neural networks.

Kalmar has attempted to combine reinforcement learning and robot learning problems (Z. Kalmar, C. Szepesvari, and A. Lorincz, Module-based Reinforcement Learning: Experiments with a Real Robot, Machine Learning, vol. 31, pp. 55-85, 1998).

Naito described above implemented a controller for an autonomous robot on evolutionary hardware using a genetic algorithm.

In addition, Koza, mentioned above, has written a simulation program that avoids obstacles using genetic programming (JR Koza, Obstacle-Avoiding Robot, in Genetic Programming II: Automatic discovery of reusable programs , MIT Press, pp. 365). -376, 1994.), Wilson wrote a robotic simulation program to find a destination in grid space using genetic programming (MS Wilson et al, Evolving hierarchical Robot Behaviors, Robotics and Autonomous Systems , pp. 215-230). , 1997).

Lee also evolved the control program of a robot that finds an object and pushes it to its destination (WP Lee, J. Hallam, and HH Lund, Applying Genetic Programming to Evolve Behavior Primitives and Arbitrators for Mobile Robots, IEEE). International Conference on Evolutionary Computation , pp. 501-506, 1997).

On the other hand, when looking at the hardware evolution mode, most of the above-described conventional research results are not implemented on the hardware, the experiment was performed on the simulation.

Circuit switching of objects may be done in simulation or in hardware. The former is called the Extrinsic mode and the latter is the Intrinsic mode (P. Layzell, Reducing Hardware Evolution's Dependency on FPGAs, Proceedings of the Seventh International Conference on Microelectronics for Neural, Fuzzy and Bio-inspired Systems 1999 , pp. 171-178, 1999).

The extrinsic mode has the following disadvantages. First, the extrinsic mode requires longer time to get results than the intrinsic mode because of the computational costs in modeling and quantification. Second, there is no guarantee that circuits evolved in extrinsic mode will work in real hardware. Third, when using the intrinsic mode, factors that were not considered in advance may be reflected, whereas in the extrinsic mode, the limitations of the factors considered in the simulator implementation process cannot be overcome. Therefore, many researchers conduct research using intrinsic mode (A. Thompsom, An Evolved Circuit, Intrinsic in Silicon, Entwined with Physics, First International Conference on Evolvable Systems 1996 , pp. 390-405 , 1996. etc).

However, research using extrinsic mode is not unnecessary. First, in the extrinsic mode, the current and voltage measurements of each trunk and node can be accurately measured. Second, in the extrinsic mode, there are no restrictions on the types of blocks used. It is allowed to decorate. Third, in intrinsic mode, various hardware for evolution are not provided.

In the present invention, circuit evolution is performed in an intrinsic mode. Those skilled in the art can use both high and low language for the representation of the circuit. Describing circuits using high-level languages requires the creation of a configuration bit stream to the underlying hardware, but with the advantage of reducing the size of the search space.

The method of directly using the configuration bit stream can be seen in the study of de Garis. In his research, he constructs and evolves objects with basic gate-to-gate connections such as AND, OR, and NOT (H. deGaris, Evolvable Hardware: Genetic Programming of a Darwin Machine, Artificial Neural Nets and Genetic Algorithms , pp. 441 -449, 1993).

Using a low-level language, such as de Garis's research, has the advantage of being able to express objects directly without any transformation.

However, there are several approaches to using the configuration bit stream as an evolutionary entity. First, the size of an object is so large that current techniques cannot efficiently perform evolutionary operations. Second, most of the bit streams that make up an object represent circuits that are not valid. Will be configured.

Evolutionary hardware, on the other hand, is being applied to adaptive systems, fault-tolerant systems, and design automation.

Here, the adaptive system refers to a system that can adapt to new changes by dynamically reconfiguring the hardware structure by learning environmental changes or mission changes.

Stoica proposes an automatic reconstruction method through evolution on PTA (Programmable Transfer Array). The Stoica group is studying how to increase the lifespan of interstellar probes through adaptation, and deals with dynamic reconfiguration of hardware-specific circuits to improve computational power (A. Stoica, D. Keymeulen, C. Lazaro, W. Li, K. Hayworth, and R. Tawel, Toward On-board Synthesis and Adaptation of Electronic Functions: An Evolvable Hardware Approach, Proceedings of 1999 Aerospace Conference , pp. 351-357, 1999).

Kajitani implemented the prosthetic control circuit using evolutionary hardware. The adaptive Kajitani group's control circuits outperform neural network-based control circuits and significantly reduce the time between user and handicap (I. Kajitani, T. Hoshino, D. Nishikawa, H. Yokoi, S. Nakaya, T. Yamauchi, T. Inuo, N. Kajihara, M. Iwata, D. Keymeulen, and T. Higuchi, A gate-level EHW chip: Implemeting GA Operations and Reconfigurable Hardware on a single LSI, Second International Conference on Evolvable Systems 1998 , pp. 1-12, 1998).

Tanaka has implemented data compression and decompression chips for electrophotographic printers using evolution hardware (T. Hikage, H. Hemmi, and K. Shimohara, Compression of Evolutionary Methods for Smoother Evolution, Second International) Conference on Evolvable Systems 1998 , pp. 115-124, 1998).

In addition, research on the fault-tolerant system based on the evolutionary hardware to secure the spare block to replace the defective block.

Ortega-sanchez maintains a copy of each block of hardware that contains the configuration information for that block so that it can be replaced if a fault occurs (C. Ortega-Sanchez, and A. Tyrell, Fault-). Tolerant Systems: The Way Biology Does It, Proceedings of the 23rd Euromicro conference , pp. 146-151, 1997).

Macias presents the possibility of a fault-tolerant system in his Processing Integrated Grid (PIG) architecture. Each cell in the PIG structure may be reconfigured as one of the function functions of neighboring cells. This suggests the possibility of a system that transfers the function of a defective block to surrounding cells (NJ Macias, The PIG Paradigm: The Design and Use of a Massively Parallel Fine Grained Self-Reconfigurable Infinitely Scalable Architecture, Proceedings of the First NASA). / DoD Workshop on Evolvable Hardware , pp. 175-180, 1999).

Moreno has proposed Field Programmable System On a Chip (FIPOC) with self-repair capabilities (JM Moreno, J. Madrenas, R. Kielbik, J.Faura, and JM Insenser, Realization of Self-Repairing and Evolvable Hardware). Structures by Means of Implicit Self-Configuration, Proceedings of the First NASA / DoD Workshop on Evolvable Hardware , pp. 182-187, 1999).

There is also a lot of research into using evolutionary hardware for design automation.

Thompson automatically designed the electronic circuits under the control of genetic algorithms. He succeeded in automatically designing the circuit that separates the signal from the signal.

JFMiller designed two-bit adders and two-bit multipliers automatically (JF Miller, and P. Thompson, Discovering Novel Digital Circuits using Evolutionary Techniques, IEE Half-day Colloquium on Evolvable Hardware Systems , 1998).

In addition, there has been an interesting research movement in the implementation of evolutionary hardware. Some researchers have begun to study evolutionary hardware implementations of block structures that can select one of the predefined functional functions at runtime and reconstruct them into those functions.

In addition, Miller (NLMiller) is MDL (Multiply-Divide) that can perform functions such as AND, OR, NAND, NOR, NOT, MUX, ADD, SUB, MULTIPLY, DIVISION in RA-FPGA (Reconfigurable Arithmetic FPGA) structure (Logic) is a function block (NL Miller, and SF Quigley, A Reconfigurable Integrated Circuit for High Performance Computer Arithmetic, IEEE International Symposium on Circuits & Systems , 1999).

By using the MDL structure described above, Miller has opened up the possibility of a hardware structure that can have high computational performance without sacrificing flexibility.

Macias proposes a hardware block structure that can be reconfigured by selecting a predefined function function from the processing integrated grid (PIG) structure proposed by him and suggests the possibility of hardware with both fault tolerance and adaptability. (NJ Macias, The PIG Paradigm: The Design and Use of a Massively Parallel Fine Grained Self-Reconfigurable Infinitely Scalable Architecture, Proceedings of the First NASA / DoD Workshop on Evolvable Hardware , pp. 175-180, 1999).

And evolutionary hardware is attracting attention as a new tool that can provide both the computational power of an ASIC and the flexibility of a general-purpose processor.

This is because the dynamic reconstruction capability of the evolutionary hardware enables the computational performance to be improved by expressing the computational circuit specialized for the process on the hardware in accordance with the replacement of the current computational process. To increase computational performance through this approach, partitioning of the processes to be performed in hardware is essential.

However, the related arts have a problem in that a problem such as autonomous mobile robot control cannot be implemented in limited hardware due to capacity problems.

Genetic programming using process partitioning and a system using the same according to the present invention introduce a method called context switching for the partitioning of the process. This method is a way of getting ideas from gene programming.

The context switching method is a method based on the excellent expression ability of gene programming, which makes it possible to implement a task for solving problems in limited hardware in a problem such as autonomous mobile robot control due to capacity problem.

The method uses gene programming to evolve a gene tree that functions like the target process. Since the subtree of the gene tree means a subprocess of the entire process, a subtree to be expressed in hardware is selected. The process chosen as the hardware computation circuit is separated from other subtrees using special nodes.

Each function node in the subtree represents a function block in the evolutionary hardware, and the connection between the function nodes means routing between the function blocks.

When the selection of the subtree is completed, the subtree is converted into hardware representation information and represented on the evolutionary hardware at the corresponding calculation time.

Through the context switching method described above, firstly, it is possible to implement a new approach that combines the advantages of ASICs and general-purpose processors, and secondly, in the hardware representation of the genetic program, it is possible to select an object regardless of the given hardware size. It makes it possible to express in hardware.

In order to verify the efficiency of the method, the present invention applies to the autonomous mobile robot control problem as an embodiment.

The contents of the above embodiment are as follows. There are two robots (KHEPERA), obstacles, destinations, and an environment composed of pushed objects. Two robots aim to collaborate and move objects to their destination while avoiding obstacles in an obstacle environment. It is an object of this embodiment to implement a control circuit which enables the achievement of a goal using evolutionary operations including the context switching method.

1 is an example of a depth-related cross operation,

Figure 2 is an example of the evolution process using the fitness conversion step,

3A is a process division diagram through the context switching module;

3B is a process replacement diagram of a subtree represented by a hardware library,

4 shows a process division process in the context switching step,

5 is an example of an individual robot control program using genetic programming,

6 is a structural diagram of a function node expression block,

7 is a system diagram of an autonomous mobile robot according to an embodiment of the present invention;

8 is a graph of the fitness of the optimal individual as a result of the above embodiment;

9 is an average goodness of fit graph,

10 is a graph of proximity recovery to the target.

<Explanation of Signs of Major Parts of Drawings>

10: evolutionary unit 20: partitioning unit 30: hardware library

40: robot control unit 50: communication unit 60: evolutionary hardware unit

Genetic programming using process partitioning according to the present invention is a gene programming method for performing an evolutionary operation including a process partitioning process for the implementation of evolutionary hardware,

A fitness switching step for improving a fitness operator used in the evolutionary operation, and a context switching step of automatically dividing and converting the process to perform the evolutionary operation. .

The relevance conversion step may include: defining two or more basic actions for the problem space to be searched, defining micro actions for problem solving through combining the basic actions, and a set of goodness of fit for the micro actions. Defining a combination of micro actions for a target action for solving the problem, defining a gene program having two or more subtrees, evolving the combination of micro actions according to the gene program Characterized in that it comprises a.

The context switching step may include analyzing available hardware resources, selecting a function node and an end node to form a tree, selecting a condition node used as a root node of a subtree, and selecting the function node and an end node. Evolving the tree by using; dividing the evolved tree into subtrees of a size that can be represented in the hardware; converting the divided subtree into a form for hardware representation; Constructing a hardware circuit including a tree and a condition node for the subtree, and expressing the hardware circuit in the hardware; and calculating a final result by adding mid-results. It features.

The evolutionary operation includes a depth-dependent crossover that selects an intersection point based on the intersection selection probability at every depth, and the intersection selection probability is fixed in advance, and then the same probability value is continuously applied. Characterized in that.

In addition, a system using a gene programming method using process partitioning according to the present invention is a system using gene programming with a process partitioning strategy for an evolutionary hardware application.

And a context conversion module for automatically dividing and converting the process to perform the evolutionary operation.

The context switching module comprises an evolver that expresses a given process using a predefined function node and an end node, and receives the evolved tree from the evolved part into subprocesses and implements the partitioning result in hardware. In a possible form, a divider constituting a hardware library is included, wherein the hardware library stores information for hardware implementation of the divided subprocess.

Therefore, the gene programming method using the process partitioning and the system using the same according to the present invention proposes and implements a method for using the gene programming as a dynamic reconfiguration method of evolutionary hardware.

Evolvable hardware is reconfigurable hardware that is designed to automatically adapt to changes in the environment or changes in tasks under the control of evolutionary operations and performs dynamic reconfiguration.

Genetic programming is a stochastic search technique suitable for inferential learning tasks. Genetic programming described above is an idea from genetic phenomena occurring in nature, and through the process of increasing the performance of an individual using a genetic operator such as a cross operator or a mutation operator to find an individual performing a desired function.

The gene programming expresses a program using a tree structure or an expression structure similar to the S-expression of LISP. Therefore, the genetic program has the expressive ability to express various functional programs.

The process partitioning refers to a process of rearranging the processes that do not require parallel operation and then rearranging them in rows of subprocesses according to the calculation order. The sequence analysis results of the subprocesses are used to generate hardware calculation circuits specialized for each subprocess.

The above object refers to a program created to search for a program having a desired function. The set of individuals described above is called a population.

Genetic programming described above finds a desired solution through the following process.

1. Randomly generate populations representing candidate solutions.

2, each solution is given a goodness of fit indicating the degree of proximity between the desired solution and the candidate solution.

3. Select the operands of the genetic operation based on the goodness of fit for each solution.

4. Construct a new population using the selected individuals and assign new fitness.

When solving the problem using the genetic programming described above, it is a matter of selecting a function node and an end node to be determined at the beginning stage.

The function node is a kind of node forming a tree of gene programming, and is composed of a function such as binary operation, arithmetic operation, special operation, and the like applied to a gene programming system.

The end node is composed of an input and a constant of the entity or an output of the entity.

In connection with the autonomous mobile robot which is an embodiment of the present invention, the function node is a node in charge of the interpretation of sensor information, and the end node is a node representing the behavior of the robot.

And crossover operators are important in evolutionary operations. The improvement of the fitness of the individual is believed to be mainly due to the cross operator (S. Luke, and L. Spector, A Comparison of Crossover and Mutation in Genetic Programming, Proceedings of Genetic Programming 1997 , pp. 240-248, 1997).

The crossover operator is a basic operator of gene programming. After selecting two entities to act as parent entities, a new entity is created by selecting and exchanging subtrees from each entity.

The crossover operator is an essential operator in gene programming, but a common crossover operation has a problem in that the intersection point is randomly selected regardless of the structure of the tree.

That is, in the case of randomly selecting intersections, firstly, the schema of the tree may be damaged, and secondly, the intersection may be concentrated near the end node due to the numerical relationship of nodes (T. Ito, H. Iba, and S. Sato, A Self-Tuning Mechanism for Depth-Dependent Crossover, in Advances in Genetic Programming 3 , MIT Press, pp. 377-399, 1999).

In order to overcome the above problems, the present invention uses a depth-dependent crossover.

1 shows an example of a depth-related crossover operation. In the depth-related cross operation, the intersection selection probability is given at every depth of the tree, and the intersection point is selected based on this probability.

In the present invention, a method of applying the same probability value after fixing the intersection selection probability in advance is adopted.

In the above method, the tree may continue to grow according to the selection of the subtree to be replaced, and the growth of the tree may not be limited. However, in the present invention, the maximum depth of the tree is limited. This is because abandoning the growth of the tree may cause inefficiency due to the addition of meaningless nodes.

That is, when a subtree exceeding the limit depth is generated, a method of maintaining the limit depth by selecting one of the subtrees and exchanging with the subtree exceeding the limit depth is selected. At this time, the subtrees to be exchanged are randomly selected.

The mutation operator is an operator that selects one of the nodes constituting the individual and replaces it with another node. The mutation operator is an operator that is critical for finding hidden nodes. And, if the population is small, mutation operators may be more efficient.

Therefore, in the present invention, the mutation operator is implemented as an operator that selects one node among the nodes constituting the individual and converts the same into another kind of node.

On the other hand, the goodness-of-fit operator is an operator for representing the object and the proximity of the object in the problem space (W. Banzhaf, P. Nordin, RE Keller, and FD Francone, in Genetic Programming an Introduction , Morgan Kaufmann Publishers, Inc.). 1998).

The design of the goodness-of-fit calculation serves to direct the search process as the only part of the researcher's influence in the solution search. The design of the goodness-of-fit calculation may vary, but problems such as robot control, which is an embodiment of the present invention, typically require two or more goodness of fit evaluations. However, the current method of genetic programming does not provide a suitable way to solve this problem.

In the genetic programming method using the process partitioning and the system using the same according to the present invention, a fitness switching step is introduced to solve such a complicated kind of problem.

The fitness switching step is a methodology for solving a problem based on a gradual learning process and finds a solution to the problem through the following process.

1. Define the basic actions for the problem space you want to explore. This process corresponds to the selection of the end node.

2. Define micro action B _i for problem solving through the combination of the above basic actions. That is, micro behavior B _i is defined as B = {B ₁ , B ₂ , ..., B _n }.

3. Define a set F of goodness-of-fit f _i for the micro action B _i . That is, the set F of the goodness of fit is defined as F = {f ₁ , f ₂ , ... f _n }.

4. Define the combination S _t of the micro actions to achieve the target action, i.e. the combination S _t of the micro actions is defined as follows.

S _t = S _t-1 (+) B _t

The (+) operator means a concatenation operator and S ₀ has a null value.

The corresponding fit function combination is defined as follows.

F _t = F _t-1 (+) f _t

5. Define a gene program A with n subtrees. Below the root node is the subtree A = (A ¹ , A ² , ..., A ⁿ ).

6. Evolve S _t in turn.

Find all the goodness-of-fit functions f _i for each microbehavior, and then use them to find the overall goodness of fit.

Figure 2 is an example of the evolution process using the above fitness conversion step. Each subtree A ⁱ is a tree for evolving microbehavior B _i .

Genetic programming using process partitioning and a system using the same according to the present invention introduce a context switching method for performing the above evolutionary operation.

The context switching method is a method for automatically dividing a process for adaptive system application of evolutionary hardware.

If a process intended to run on evolutionary hardware does not require parallel operations, the process can be divided into a set of processes that are performed sequentially. The context switching method is based on this observation.

The method divides a large process into a set of small processes and then executes the divided subprocesses in the order of calculation.

The context switching module for the execution includes an evolver, a divider, and a hardware library.

3A illustrates a process division process through the context switching module.

The evolution unit 10 evolves a tree representing a given process using a function node and an end node.

The division unit 20 receives the tree evolved by the evolution unit, divides the tree into sub-processes, converts the division result into a hardware-implementable form, and delivers the result to a hardware library.

The hardware library 30 stores gate types and routing information for hardware implementation of the subprocess.

3B is a process replacement diagram of a subtree represented by a hardware library. P _i denotes the i th process stored in the hardware library, and r _i denotes the result of the i th process. Hardware representation information stored in the hardware library of the i-th process is implemented in the evolutionary hardware, receives an environment input, and stores the calculation result in r _i .

4 sequentially illustrates a process division process in the context switching step. In FIG. 4A, a rectangle indicated by a dotted line means a given hardware resource. The whole tree is divided into subtrees of a size that can be represented in hardware, and the final result is formed by adding the intermediate calculation results.

As shown in FIG. 4, the context switching method is executed through the following process.

1. First, analyze available hardware resources.

2. Select the function nodes F = {f ₁ , f ₂ , ..., f _n } and the terminal nodes T = {t ₁ , t ₂ , ..., t _n } to form the tree.

3. Select the conditional node C = {c ₁ , c ₂ , ..., c _j }.

The conditional node is a special kind of node which is an element constituting a function node but is not selected as an operand of a cross operation and a mutation operation.

The condition node is used as a flag. In addition, the condition node is used as the root node of the subtree, and only when the value of the condition node is satisfied, the hardware circuit is changed to a circuit representing the next sub process.

4. Evolve the entity-the gene tree-using the function and end nodes defined above.

5. The evolved tree is divided into subtrees of a size that can be represented in hardware. The splitting operation does not proceed to the node of an atomic unit, and is stopped if it is a size that can be represented in a given hardware (step (b) of FIG. 4).

6. Convert the partitioned subtree into a form for hardware representation.

The type of the conversion gate and the routing information between the function blocks are stored in a hardware library.

7. The partitioned subtree and the condition node C _i for the tree are combined to form a hardware circuit, and then expressed in hardware. The calculation result of the circuit is stored in a register or an external memory (step (c) of FIG. 4).

8. The final result is calculated by adding the intermediate results (Mid-Result) (step (d) of FIG. 4).

The second to fourth processes are performed in an evolver, and the fifth process is performed in a divider.

The context switching method is a process of dividing a tree into several small subtrees, and then obtaining a result by selecting and exchanging a tree to be implemented in hardware.

Genetic programming can be a fairly reasonable choice when segmenting a process using such context switching.

The tree-structured entity of gene programming does not require Data Flow Analysis for process partitioning.

Therefore, in gene programming, the process of dividing a process can be completed only by selecting a subtree.

In order to apply a specific process to the context switching, relationship complexity and dwelling period must be considered.

The relationship complexity refers to the complexity of the relationship between subprocesses. For example, the relationship complexity increases when there is a process in a recursive call relationship or a cycle exists.

The residence period means the time taken to finish the subprocess. If the subprocess includes a loop, the residence period is long.

The relationship complexity and residence period serve as a criterion for applying context switching to a process.

Here, the residence period is a more important factor in the evolutionary hardware implementation. If the residence period is less than a certain value, there may be a performance degradation due to the exchange time of the context on the hardware. In such cases, it may be efficient to execute the task using a general purpose processor.

As an embodiment according to the present invention, an autonomous mobile robot is characterized by low relationship complexity and long residence period and can be easily applied to evolutionary hardware.

Example (autonomous robot control)

Conventional robots are suited for repetitive and unchanging missions. These kinds of robots are programmed with very detailed information about the tasks to be performed and the environment to be performed.

However, in the case of a mobile robot, the data is received with less reliable data far from the well-defined environmental information of the conventional robot.

Therefore, a robot that must perform a task in a dynamic environment needs to have some judgment ability to complete the task.

An embodiment according to the present invention is an advanced method and system in which two robots push an object to a destination while avoiding obstacles.

The robot used in the above embodiment senses the environment using one line vision sensor and eight infrared sensors.

5 shows an example of an individual robot control program using gene programming. Each path from the root node to the end node in the tree of FIG. 5 represents a sensor information analysis path.

The function node of the above embodiment consists of the following seven elements.

1.IF-OBJ: If the robot finds an obstacle,

If not found, the subtree on the right is executed.

2. IF-GOAL: If the robot finds a destination, it executes the subtree on the left, and if it does not find it, it executes the subtree on the right.

3. IF-FORWARD: When the robot finds an object and finds its destination, it has a function to determine whether it is possible to push the object. The subtree on the left is executed when arranged in the order of destination-light source-robot, and the subtree on the right when not arranged.

4 ~ 7. IF-OBS1, OBS2, OBS3, OBS4: Determine whether there is a set of obstacles among 4 IR sensor sets. If there is an obstacle, the subtree on the left is executed. If not, the subtree on the right is executed.

The end node of the above embodiment is composed of the following seven elements.

1. MF: Advances the robot in the direction observed by the line vision sensor.

2. MTL: Rotate the robot counterclockwise and move forward.

3. MTR: Rotate the robot clockwise and move it forward.

4. MB: Back up the robot.

5. TL: Rotate the robot counterclockwise.

6. TR: Rotate the robot clockwise.

7. RANDOM: Randomly select one of the six exercises above.

Referring to the object of FIG. 5 as an example, if the root node interprets the sensor information and finds a push object, the control right is transferred to the left node if the right node is not found. The subtree that receives the control right interprets the sensor information using the function of the node to determine which path is valid.

When the process continues to reach the end node, the information possessed by the end node corresponds to the motion information of the robot.

In addition, the micro action required for the autonomous mobile robot control, which is an embodiment of the present invention, is heading and homing.

The heading refers to a process of finding an object to be pushed toward a destination. The goodness-of-fit function defined for the heading is shown in Equation 3.

F _new = F _old + w ₁ xn (collisions) + w ₂ xn (steps)

(In the above equation, the symbol "x" means multiplication.

The homing is a process of finding an object and transporting it to a destination. The fitness function defined for the homing behavior is expressed by Equation 4.

F _new = F _old + w ₁ xn (miss) + w ₂ xn (steps) + w ₃ x vision

And if the suitability is given to each individual, the individual to make up the next generation should be selected. This role is the selection operator. That is, the selection operator is an operator that selects an entity to construct the next generation based on the goodness-of-fit value assigned to each entity.

The operator used in the present invention is a goodness-of-fit selection operator.

In the fitness-proportional selection operation, each individual has a survival probability in proportion to the fitness. The survival probability is given based on Equation 5.

( Q f _i means f ₁ + f ₂ + ... + f _n .)

Enough depth of objects must be ensured to ensure a path for sensor information interpretation. However, in the ongoing execution of genetic programming, nodes may be added that are not related to valid calculations. In other words, this phenomenon is called an intron problem. Many opinions exist about the value of the intron problem, but in terms of computational cost, it is not a positive factor for hardware-implemented gene programming.

Therefore, in the present invention, individual growth beyond a certain depth is prevented to prevent excessive introns.

And, there are two kinds of blocks on evolutionary hardware to facilitate context switching according to the present invention.

The node 1 block is a node for representing a function node, and the node 2 block is a node for representing an end node. Node 1 and Node 2 are designed to represent all elements of the function node and all elements of the end node.

6 is a block representing a function node as the structure of node 1. FIG.

In the initialization state, all Node1 blocks on the hardware show the same gate distribution and inter-gate connection as shown in FIG.

However, when the context of the program to be implemented is determined by executing the program, the function node is represented by changing the type of gate marked with "*" according to the information of the context. End nodes express the desired functionality in the same way as function nodes.

7 is a system diagram of an autonomous mobile robot according to an embodiment of the present invention.

The whole system consists of part 1 (KHEPERA part) and part 2 (host PC part).

The part 1 includes a sensor and a motor controller for inputting environmental information through a sensor. Motor control is achieved by receiving motion information from part 2 and interpreting it.

The part 2 includes a robot controller 40, a communication unit 50, and an evolution hardware unit 60.

The main function of the robot control unit 40 is to convert the motion information into a form that can be interpreted by the robot, convert it to a form that can interpret the sensor information of the robot, and calculate a fitness using the sensor information. In addition, it serves to generate the motion information of the robot for the initialization state.

The communication unit 50 transmits the information sent by the robot control unit to the evolution hardware, and sends the motion information output by the evolution hardware unit to the robot control unit.

In addition, it generates a fitness graph by receiving the fitness of the robot controller, and plays the role of storing the object information indicating the best fitness.

The evolution hardware unit 60 is located on the xc6216 FPGA of XILINX Corporation. The xc6216 FPGA is a H.O.T. It is installed on the board and communicates with the program through the PCI slot of the host PC.

The robot control program is included in the FPGA of the evolutionary hardware part, and it is responsible for the interpretation of the sensor information and the determination of the motion information.

8, 9 and 10 show graphs of the results of an embodiment according to the present invention.

8 is a graph of the fitness of the optimal individual. The vertical axis of the graph represents the goodness of fit and the horizontal axis represents the generation. Looking at the graph of FIG. 8, it can be seen that the suitability is lowered since the 20th generation. However, although 20 to 60 generations have deviations in the goodness of fit, it can be seen that after 70 generations, they are stabilized.

9 is an average goodness of fit graph. Although there is considerable variation, it can be seen that the fitness value is steadily decreasing. 9, it can be seen that individuals of a generation have some degree of difference, but all of them are learning to some extent.

10 is a close-up graph to the target. As can be seen from the graph, as the generation progresses, the number of approaches to the destination increases.

As mentioned above, although this invention was demonstrated to the Example, this invention is not limited to this. Although robot control has been described as an embodiment of the present invention, there are various industrial applications of the evolutionary hardware to which the present invention is applied. The present invention can be applied to a field in which various algorithms for improving performance of operations, such as a control processor of planetary exploration, a processor for data compression in wired and wireless communication, must be used in sequence, or an area that needs to adapt independently to environmental changes. Can be.

Genetic programming using process partitioning and a system using the same according to the present invention divides a complex process into small and simple subprocesses using context switching and fitness switching together, and then implements the partitioning result in hardware. By converting to possible forms, executing subprocesses according to the calculation order, and combining the intermediate calculation results to form the final result, it is possible to implement an adaptation system that can adapt to environmental or mission changes.

That is, the harvest which can be obtained by this invention is as follows.

First, it proposes a way to express the tasks that could not be expressed in a given hardware due to capacity problems. Second, it is possible to implement a circuit that can adapt to environmental changes by adjusting control circuits.

In addition, it proposes a method for developing an adaptive system using the dynamic reconfiguration capability of evolutionary hardware.

Claims (6)

In gene programming using process division, Analyzing available hardware resources; Selecting a function node and an end node to form a tree; Selecting a condition node used as a root node of a subtree; Evolving the tree using the function and end nodes; Dividing the evolved tree into subtrees of a size that can be represented in the hardware; Converting the partitioned subtree into a form for hardware representation; Constructing a hardware circuit including the divided subtree and a condition node for the subtree, and expressing the hardware circuit in the hardware; Calculating the final result by adding the intermediate results; A context switching step of automatically dividing and converting a process; Gene programming method using a process partition comprising a. The method of claim 1, The context switching step, Defining at least two basic behaviors for the problem space to explore; Defining a micro behavior for problem solving through combining the basic behavior; Defining a set of fitness for the microbehavior; Defining a combination of micro actions for a target action for solving the problem; Defining a gene program having two or more subtrees; Evolving the combination of micro-actions according to the genetic program; and further comprising a fitness conversion step of improving a fitness operator. The method according to claim 1 or 2, The context switching step includes a depth-dependent crossover that selects the intersection point based on the probability of selecting the intersection point at every depth, and the probability of selecting the intersection point is fixed after applying the same probability value. Genetic programming using process partitioning characterized in that. As a system of using genetic programming using process partitioning, A context switching module for automatically dividing and converting a process; Genetic programming using system using a process partition comprising a. The method of claim 4, wherein The context switching module, An evolver for expressing a given process using a predefined function node and an end node; A process division comprising: a divider configured to receive a tree evolved by the evolution unit as an input and divide the sub-process into a sub-process, and express the division result in a form that can be implemented by hardware to form a hardware library. Genetic programming using system. The method of claim 5, And the hardware library stores information for hardware implementation of the divided sub-process.