Martin V. Butz, Oliver Herbort, Joachim Hoffmann, Andrea Kiesel, Alexandra Lenhard
MindRACES: FP6 – 511931
MindRACES is an interdisciplinary research project that integrates seven European labs with expertise in cognitive science, cognitive psychology, and artificial intelligence. The project focuses on the notion of anticipatory behavior in cognitive, adaptive, embodied learning systems. Anticipatory behavior may be defined as a mechanism that uses (predictive) knowledge about the future to improve current behavior. Such behavior includes (model) predictive control mechanisms, decision making, planning, filtering, as well as attention. A cognitive embodied system is an adaptive system that interacts with a simulated or real environment by the means of a defined, artificial body, such as a (simulated) robot.
Two major (possibly interacting) mechanisms are necessary for efficient anticipatory behavior: (1) A predictive model needs to be available or needs to be learned in order to make reliable predictions about the future. (2) The knowledge about the future needs to be exploited (selectively) to optimize current behavior.
The involvement of the Würzburg group stems from Hoffmann’s proposition of the cognitive behavioral and learning mechanism of Anticipatory Behavior Control (Hoffmann, 1993; Hoffmann:2003), the subsequent creation of Stolzmann’s Anticipatory Classifier System (Stolzmann, 1997, Stolzmann, 1998), and the successive studies and improvements adding generalization and an independent reinforcement component (Stolzmann, Butz, Hoffmann, & Goldberg, 2000; Butz, 2002; Butz, Goldberg, & Stolzmann, 2002, Butz & Hoffmann, 2002, Butz & Goldberg, 2003). The premise of ABC is that any voluntary behavior is initiated and controlled by a representation of the to-be attained (sensory) properties. The difference between desired properties and current state triggers and “guides” the behavior towards the goal. The ACS system is a rule-based system that learns explicit condition-action-effect (C-A-E) rules that enables the realization of anticipatory behavior as proposed in the ABC theory.
The current involvement in the MindRACES project focuses on the creation of anticipatory systems that developmentally learn (1) predictive models in dynamic environments, (2) the anticipatory control of a robot arm, as well as (3) the combination of the two capabilities in an autonomous anticipatory cognitive embodied system.
In order to be able to successfully pursue goals, any organism needs knowledge of the possible consequences of actions and under what conditions these action-effect relations hold. As the world is subject to continuous change, the organism must be able to extract this information by exploring the world autonomously. Additionally, the organism needs to be able to distinguish environmental dynamics caused by own actions from other dynamics in the environment.
The goal of the MindRACES project is to design competent computational models of the organization of such anticipatory, goal-oriented behavior. Thereby, it is investigated how an abstract goal representation can be transformed into actual motor behavior. The artificial agent learns inverse models that map the current state and the desired (possibly partially defined) goal into the behavior that promises to achieve this goal maximally efficiently (with least effort). It is aimed for a cascade of such models that successively decode abstract goal states into currently attainable (sub-)goals – eventually causing the execution of the currently most appropriate motor action. Similarly, we create forward models that are used to predict the environmental dynamics improving the agent’s sensory processing, making it more noise robust, and enabling it to bridge possible time delays in sensory processing.
To achieve this endeavor, we will be using mechanisms derived from ACS, the related but more flexible XCS system (Wilson, 1995; Butz, Kovacs, Lanzi, Wilson, 2004; Butz, 2006) as well as several neural network approaches including fully-connected associative (Hebb-based) networks, the self-organizing Neural GAS architecture (Martinetz, Berkovitch, & Schulten, 1993), and long-short term memory mechanisms (LSTM) (Hochreiter & Schmidhuber, 1997).
Anticipatory behavior appears ubiquitous in living systems. The MindRACES project strives to not only understand the actual mechanisms involved in these systems but also to exploit these mechanisms to create more competent and flexible, real-time, cognitive, embodied artificial, adaptive agents. Würzburg contributes to this endeavor with its profound knowledge in cognitive psychology as well as its expertise in various symbolic and sub-symbolic, adaptive learning systems including (anticipatory) learning classifier systems as well as various neural network architectures.