James M. Hereford

Using the Particle Swarm Optimization Algorithm for Robotic Search Applications

This paper describes the experimental results of using the particle swarm optimization (PSO) algorithm to control a suite of robots. In our approach, each bot is one particle in the PSO; each particle/bot makes measurements, updates its own position and velocity, updates its own personal best measurement (pbest) and personal best location (if necessary), and broadcasts to the other bots if it has found a global best measurement/position. We built three bots and tested the algorithm by letting the bots find the brightest spot of light in the room. The tests show that using the PSO to control a swarm can successfully find the target, even in the presence of obstacles

Robust sensor systems using evolvable hardware

This paper describes a system that is robust with respect to sensor failure. The system utilizes multiple sensor inputs (three in this case) connected to a programmable device (FPAA) that averages the outputs from the sensors. The programmable device is programmed using evolvable hardware (EHW) techniques. If one or more of the input sensors fails, then the controller detects the failure and initiates a reprogramming of the circuit. The system then continues to operate with a reduced number of sensors. The failure detection is accomplished by comparing the actual system output with a Kalman filter estimate of the output. If the actual output and the filter estimate differ by amount greater than the system uncertainty, then a failure is noted. The system is robust to several different failure modes: sensor fails as open circuit, sensor fails as short circuit, multiple sensors fail, FPAA input amplifier failure. The experimental setup is described as well as results using simulated inputs (individual voltage sources) and actual temperature sensors (PRTDs). The paper also discusses some interesting results as well as issues that must be overcome to expand the system.

Bio-Inspired Search Strategies for Robot Swarms

Our goal is as follows: build a suite/swarm of (very) small robots that can search a room for a “target”. We envision that the robots will be about the size of a quarter dollar, or smaller, and have a sensor or sensors that “sniff” out the desired target. For example, the target could be a bomb and the robot sensors would be a chemical detectors that can distinguish the bomb from its surroundings. Or the target could be a radiation leak and the sensors would be radiation detectors. In each search scenario, we assume that the target gives off a diffuse residue that can be detected with a sensor. It is not very efficient to have the suite of robots looking randomly around the room hoping to “get lucky” and find the target. There needs to be some way to coordinate the movements of the many robots. There needs to be an algorithm that can guide the robots toward promising regions to search while not getting distracted by local variations. The search algorithm must have the following constraints:  The search algorithm should be distributed among the many robots. If the algorithm is located in one robot, then the system will fail if that robot fails.  The search algorithm should be computationally simple. The processor on each bot is small, has limited memory, and there is a limited power source (a battery) so the processor needs to be power efficient. Therefore, the processor will be a simple processor.  The algorithm needs to be scalable from one robot up to 10’s, 100’s, even 1000’s of robots. The upper limit on the number of robots will be set by the communication links among the robots; there needs to be a way to share information among the robots without requiring lots of communication traffic.  The search algorithm must allow for contiguous movement of the robots. This chapter will describe two search strategies for robot swarms that are based on biological systems. The first search strategy is based on the flocking behavior of birds and fishes. This flocking behavior is the inspiration behind the Particle Swarm Optimization (PSO) algorithm that has been used in software for many types of optimization problems. In the PSO algorithm the potential solutions, called particles, “fly” through the problem space by following some simple rules. All of the particles have a fitness value based on the value or measurement at the particle’s position and have velocities which direct the flight of the particles. The velocity of each particle is updated based on the particle’s current velocity as well as the best fitness of any particle in the group. 1

Fault-tolerant sensor systems using evolvable hardware

This paper describes a system that is robust with respect to a sensor failure. The system utilizes multiple sensor inputs (three in this case) connected to a programmable device that averages the outputs from the sensors. The programmable device is programmed using evolvable hardware techniques. If one or more of the input sensors fail, then the controller detects the failure and initiates a reprogramming of the circuit. The system then continues to operate with a reduced number of sensors. The failure detection is accomplished by comparing the actual system output with a Kalman-filter estimate of the output. If the actual output and the filter estimate differ by an amount greater than the system uncertainty, then a failure is noted. The system is robust to several different failure modes: sensor fails as open circuit, sensor fails as short circuit, partial failures, multiple sensors fail, field programmable analog array input amplifier failure. This paper describes the experimental setup as well as results using actual temperature sensors. For all failure types, the system was able to recover to within 2% of the target value

Analysis of a new swarm search algorithm based on trophallaxis

We investigate a new swarm search algorithm based on the trophallactic behavior of social insects, specifically honey bees. The new algorithm does not require any agentagent communication and does not require the agents to know position information. The agents, or bots, cluster together near peaks in the search space based on the fitness value at the locations where the agents collide. In this paper we describe the algorithm and analyze its effectiveness using a birth and death Markov chain. The analysis shows that the agents will congregate at or near the peaks, so the algorithm shows promise for using very simple robots in swarm search applications.

Analysis of BEECLUST swarm algorithm

We analyze a new swarm search algorithm based on the behavior of social insects, specifically honey bees. The new algorithm does not require any agent-agent communication and does not require the agents to know position information. The agents, or bots, cluster together near peaks in the search space based on the fitness value at the locations where the agents collide. In this paper we describe the algorithm, model the algorithm using a birth and death Markov chain, and determine the expected time for the agents/bots to cluster. We also determine the swarm size needed to complete a search in a reasonable time frame.

FlockOpt: A new swarm optimization algorithm based on collective behavior of starling birds

The aim of this paper is to introduce a new algorithm, FlockOpt, for real-parameter optimization. The proposed algorithm is inspired by a recent model of the flocking behaviour of starling birds and combines the main elements of this model with additional features from Swarm Intelligence. The results of from FlockOpt are compared to the results of generic versions of Particle Swarm Optimization, which is the closest relative of FlockOpt from the Swarm Intelligence field. The comparison shows that FlockOpt is able to beat the generic versions of particle swarm optimization in the majority of the test cases. Interesting features such as the attraction, repulsion and the alignment between members of the population make FlockOpt quite attractive for further examination.

Easily scalable algorithms for dispersing autonomous robots

This paper describes three new algorithms for dispersing a swarm of bots throughout a search space. We assume that the bots do not have a central coordinating agent and we want to have no (or few) inter-bot communications so that the algorithms can scale to large swarm sizes. We simulated the three new dispersion algorithms plus two other random-walk based dispersion algorithms on five different search spaces. Each of the five algorithms was tested with swarm sizes from three to fifty bots. For swarm sizes larger than ten, we found that the minimize-intensity algorithm, which is based on decaying signal strengths, worked best. For small swarm sizes, the dispersion algorithm based on the dispersion of gas particles performed best.

Integer-valued Particle Swarm Optimization applied to Sudoku puzzles

In this paper we develop a variation of the particle swarm optimization (PSO) algorithm that is tailored to discrete optimization problems. We focus on solving Sudoku puzzles but the ideas can be extended to other problems with discrete solutions. We compare our PSO-based algorithm to the classic PSO and to a (mu+lambda) evolutionary strategy (ES) for 50 puzzles and find that the PSO algorithms do much worse than the ES. We then consider why PSO does not do well on this type of problem.

Failure Detection for Multiple Input Systems

The overall goal of this research is to develop a system that is resistant or tolerant of sensor failures. The plan is to incorporate multiple sensors in an electronic system and then take the average of the sensors as the system output. If a sensor fails, the failure is detected, and the processing hardware will be ¿re-programmed¿ based on evolvable hardware principles to take the average of the remaining input sensors. A key aspect of this program is how to detect when one out of N sensors fails. This paper describes the use of a one-dimensional (1-D) Kalman filter to detect a sensor failure by observing the average of N sensors together. The paper discusses the failure detection algorithm and then gives results for the algorithm based on computer simulations and then actual laboratory measurements.