Separating goals from behavioral control: Implications from learning predictive modularizations

The brain may be regarded as an anticipatory machine whose behavior strongly depends on its current predictive knowledge. Behavioral decision making depends on anticipated goal states as well as on the current internal motivations of the organism. Behavioral control, on the other hand, is guided by the goals currently chosen along with additional constraints. Both, decision making and control are thus anticipatory processes. Moreover, they are mutually dependent: while action control depends on currently selected goals, goal selection depends on achievability estimates, which must be based on the system's current action control competence. An autonomous, adaptive system thus faces the challenge of learning goal representations that are suitable for both, action selection and action control. We propose that a goal processing pathway should be separated from but also strongly interact with a sensorimotor control pathway. We investigate the encoding structures expectable along these two pathways for realizing effective and flexible action decision making and control. While the goal processing pathway needs to be able to distinguish motivation-oriented relevancies for decision making, the sensorimotor pathway needs to provide control-oriented encodings. We use an evolutionary machine learning technique to investigate how important modularity may be for realizing particular sensorimotor mappings. Next, we survey the results obtained by a neural network architecture, which show that enforcing multiplicative interactions between self-organizing sensorimotor control-oriented encodings and goal-oriented interaction selection encodings enables the learning of highly flexible decision making and action control structures. Furthermore, we show that the emerging goal-oriented encodings exhibit pre-linguistic compositional structures. We conclude that for bootstrapping higher-level cognitive capabilities it may be essential on the one hand to separate sensorimotor, anticipatory, control-oriented spatial encodings from compositional, goal-oriented spaces, and on the other hand to enable bidirectional, multiplicative interactions between these two sets of spatial encodings.