Remembering the Past: recipe for the ultimate survivor? The value of multiple timescales in a recurrent neural network for self-organization of survival behavior in random versus structured environments

Thumbnail Image
Journal Title
Journal ISSN
Volume Title
It is generally thought in cognitive neuroscience that the concept of functional hierarchy - the notion that complex things can be decomposed into simpler elements and that simpler elements make up a complex system - plays an important role in the production of skilled (motor) behavior and situations that require cognitive control. According to schema theory, behavioral elements make up behavioral primitives, which can be sequenced to achieve a global goal. In robotics, there have been many different attempts to design paradigms for such behavior productions but often a distinction is made between reactive and deliberative robots. Hybrid systems incorporate both kind of behaviors, in which a higher level system controls lower level reactive layers to produce behavior (e.g. the traffic regulator concept). Since it is not really clear how such functional hierarchy is actually organized in the brain, it would be interesting to see how this functional hierarchy can self-organize. In the current thesis, a recurrent neural network model was used for such self-organization. Context units with different multiple timescales were used, to incorporate the temporal organization of behavior. The goal was to test how well such an MTRNN agent performed and what kind of behavior was shown as compared to a traffic regulator on a survival task in a day-night environment, with obstacles and food sources. Furthermore, since hybrid robots are consistent with embodied embedded cognition, it would be interesting to see what kind of role environment type plays for the behavior of the MTRNN agent. Therefore the behavior and performance of the MTRNN was tested in two different environments, varying in the amount of structure. It was found that the MTRNN agent performed worse than the other tested agents, but that performance was better in more structured environments. This implicated that the environment is an important factor but that the MTRNN agent is less suited to random environments. As for the self-organization of functional hierarchy, it did not emerge through the use of different timescales but the complexity of behavior was dependent on the right amount of food in the environment. The results indicated that in order to achieve functional hierarchy and perform well, the agent needs clear goal-directed tasks and structured environments. Keywords: functional hierarchy, skilled behavior, reactive, deliberative and hybrid robotics, embodied embedded cognition, recurrent neural network, multiple timescales
Faculteit der Sociale Wetenschappen