Open-Ended Evolution of a Circadian Rhythm

Most biological systems have some sort of adaptation to our planet’s cycle of day and night. This adaptation is a current subject of scientific research, and serves as inspiration to develop a multi-agent simulation to investigate the evolution of complexity in an open-ended evolutionary framework. In a previous work, we created a simulated world where artificial organisms evolve to synchronize with a daily cycle of light and darkness. A multi-agent, artificial life framework was used to implement these simulations. In this paper, we further develop that world, by adding caves to the environment. When in these caves, the agents will perceive a low level of light, as if it were night. This adds an extra layer of complexity to the desired behavior of the agents, as now they need to distinguish “night” from “cave”. Using the same agent structure, and the same open-ended evolution framework, we show that the agents evolve to adapt to this new environment. We also show how the agents adapt to the environment with caves, by analyzing their brains. Introduction The study of circadian clocks and similar synchronization phenomena in biological systems is a current subject of scientific research (Rand et al., 2006; Strogatz, 2004). Despite having been extensively studied, these phenomena still have much to be investigated. Our goal, however, is not to learn more about this biological process, but to use it as an inspiration to study the emergence of complex behaviors in an open-ended evolution scenario. To that end, we implement a simulation where the environment has a day and night cycle, and analyze the evolution of the agents’ behavior, and their adaptation to this cycle. In a previous paper, the authors presented some experiments done with such a scenario (Baptista and Costa, 2008), and showed that the agents do develop behaviors adapted to the daily cycle. In an effort to create a more challenging environment, we now further developed that world by adding caves. When an agent enters a cave, the light level will be the same as if it were night time. This will force the agents to evolve behaviors capable of distinguishing the two different low light conditions, adding extra complexity to the requirements for survival. Some previous work has been done with similar simulation scenarios, either by evolving neural networks (Mirolli and Parisi, 2003), or virtual CPU organisms in AVida (Beckmann et al., 2007). Our scenario can be mostly compared to that of (Mirolli and Parisi, 2003), as they also have an environment with varying light level and caves. However, they use a standard genetic algorithm to evolve the agents, whereas we use an open-ended evolution framework. Although an established definition of open-ended evolution hasn’t yet surfaced, most authors consider that one of the major requirements is the absence of an explicit fitness function. In other words, to have open-ended evolution, a system should be based on Natural Selection rather than Artificial Selection (Channon, 2000). The simulations described in this paper were implemented using the BitBang framework. One of the purposes of these simulations is to serve as a proof of concept for the model developed for the framework. Implementing a modern autonomous agent model (Russell and Norvig, 2002), this framework has roots in Artificial Life systems and Complexity Science. The simulated world is composed of entities. These can either be inanimate objects which we designate as things, or entities that have reasoning capabilities and power to perceive and affect the world—the agents. Both have traits that characterize them, such as color, size, or energy— the features. The agents communicate with, and change the environment using perceptions and actions, taking decisions using the brain. In this model, there is no definition of a simulation step, as we won’t have any type of centralized control. As such, the simulation is asynchronous. The agents will independently perceive, decide, and act. Moreover, there is no evolutionary mechanism included in the definition of the model, since evolution is implemented as an action. That is accomplished by giving the agents the capability of reproduction. Again, there is no central control bound to the process of reproduction. The agents choose when to reproduce and with what other agent to reproduce with. In addition, there is no explicit fitness function. The agents die due to lack of resources, predators, age, or any other mechanism implemented in the world. Thus, in this ECAL General Track