• No se han encontrado resultados

CARDENE IV SOLN 40-5 MG/200ML-% [nicardipine

PHENOBARBITAL TABS 30 MG

Watson, Ficici and Pollack [93] use a group of robots, running heterogeneous controllers to evolve a phototaxis behaviour in a distributed, asynchronous way. A lamp is placed in the middle of the arena, when the robot reaches the light it gains energy and moves to a random part of the arena, from there it must repeat the task to gain more energy.

When a robot has enough energy it can reproduce, it does this by broadcasting a slightly mutated version of its genome over a small local area. Other robots may take this broadcast genome as their own, completely replacing their existing genome, with a probability dependent on the receiving robot’s energy level. If it has low energy there is a higher probability that it will accept the new genome than if the robot has a high energy. This method of sharing and evolving genomes is a swarm behaviour and does not require any outside assistance or centralised controller to monitor, evolve or distribute a robot’s genome; it is all performed collectively by the swarm. It is open- ended and adaptive, so if the environment changes over time the swarm will also change in reaction.

Floreano et. al. [29] evolved simple communication in a swarm of 10 foraging robots. They demonstrated that swarms where the robots all ran the same controller, were more likely to signal to each other in the presence of food and so in general had higher fitnesses. The experiment was performed using a high fidelity simulation of the robots and the winning genome was transferred to the real, embodied robots and were found to display the same behaviour as in simulation.

Pugh and Martinoli [74] investigated the effects of genome diversity on the fitness of a simulated robot swarm running heterogeneous genomes. They compare two opti- misation algorithms. One, a genetic algorithm, used all the genomes in the population to generate the next population. In the other algorithm1 each genome generates one offspring for the next generation based on its own fitness and the fitnesses of its nearest two neighbours in the population array. In the second algorithm there is more genetic diversity in the population because all genomes are drawn from a different set of “par- ents”. The authors found that the more diverse algorithm gave higher fitnesses, even in the presence of sensor and actuator noise. When they investigated why this was the case they found that “by maintaining high diversity throughout evolution, [the more di- verse algorithm] is able to continuously discover new and better solutions and continue improving throughout the entire evolution” [74].

Heterogeneous swarms have the advantage to ESR research that the genomes can be evaluated in parallel with each other rather than sequentially, but this means that the evolutionary algorithm’s genome population must be the same as the number of robots in the swarm; with a homogeneous swarm there is no such limit. Consequently heterogeneous swarms are faster to advance from generation to generation but because population size is limited there is less genetic diversity within the genome population. Additionally, genomes are assessed by a robot’s ability to perform a collaborative task, therefore using a metric that is partly dependent on the performance of other genomes. For example, if in this experiment the GA were to somehow generate a genome which encapsulates the perfect solution but it is one of a heterogeneous population of otherwise poorly performing genomes. The perfect genome would not collect as much food as it would in a better performing population, because it still relies on the others to help it collect. An “average” solution in an “average” population may well collect more food and so be awarded a higher fitness. In a homogeneous swarm this wouldn’t be a problem because all the robots would be using the same genome.

The biological counterparts to swarm robotic systems, swarm insects, are not ge- netically identical but they also have little genetic diversity between insects in a colony. 1This algorithm is called “Particle Swarm Optimisation”. A review of the particle swarm opti-

misation algorithm is outside the scope of this work, however, readers are referred to [43] for more information. A description of the specific particle swarm optimisation algorithm used in [74] is given in [73].

With ants for example, all the ants in a colony are offspring from a very small number of queens [37], if two ants share the same parents they will, on average, share 75% of their genes [50]. Worker ants are sterile and contribute to the evolution of the colony by foraging for food and defending the ant nest from invaders, so that the colony is better able to produce the queen and male ants. Evolution and diversification of the ants occurs through new queens and males leaving the nest to breed and start new colonies. Thus, it is the ant colonies that evolve, not individual ants.

Swarm Diversity in the Experiment

In our experiment we would ideally like to use the collective evolution strategy from Watson et. al. [93]. This strategy resembles evolution in natural systems because each robot has a different genome and measures the its fitness by the robot’s ability to gather a resource, energy, whilst following the behaviour specified by its genome. Robots decide between themselves whether to reproduce with each other and create a new genome based on their comparative fitnesses. Consequently, the genetic functions of selection, crossover and mutation are all performed locally by the robots, so the evolution is diverse, asynchronous, decentralised and scalable. These are some of the properties of swarm intelligence given in section 3.2, making the evolution a swarm intelligent behaviour. With evolution that is analogous to the evolution of biological systems, the robot swarm is closer to the complex systems that Simon describes in [84]. Implementing a decentralised evolutionary strategy depends on the robots being able to pass data to each other without help from an external controller, but unfortunately the robots we have for this experiment (described later in section 6.3.4) are not able to reliably pass more than one byte of data between each other. In order to transfer genomes to and from the robots we must therefore use a computer to manage the population of genomes by collecting the fitnesses of genomes running on the robots, generating the next generation, and being responsible for distributing the new genomes to the robots. Consequently, we are not able to implement decentralised evolution because updates to the robots’ genomes must be controlled through the computer. In addition to a being single centralised controller, this also removes localisation because if the robots control the genome reproduction then a robot can only share its genome with other robots that are nearby. The external computer does not know which robots are spatially close so it cannot limit genome mutation to be between robots that are neighbours. Robots that are near to each other are just as likely to reproduce as ones which are far apart.

We approximate the evolution of Watson et. al. by using a heterogeneous swarm with a centralised controller for running the GA. This approach is much quicker to evaluate a generation than a homogeneous swarm, but if the heterogeneous swarm is unsuccessful at testing the hypothesis we can easily run the experiment with a homo- geneous swarm by changing the GA on the computer.