Learner With Body and Self

Figure 9.1: Scenario with learner having a body

After evaluating the first experiments with evolutionary connectionist systems (ECoS) we have not been satisfied with the explanatory power of these models. We decided to try another approach which later on can lead back to a neural machinery.

In figure 9.1 one can see a rough sketch of the new scenario.

Here is a list of the main assumptions:

1. General Setting: A Learner has a simple body in a room which is represented as a 2-dimensional grid. There is also another moving body. There is a sensor which measures the occupation of a grid point by a body. The measured data are formatted as a 2-dimensional matrix showing the occupied points as having distinguished values. While the movements of the non-learner body is controlled outside of the learner the learner can cause movements of the learner-body by distinguished neural states inside of his body.

2. Minimal Movements: In the basic scenario we assume as minimal movements the movement from one point to another like Left (L), Right (R), Up (U), Down (D) or do nothing NoOp (N). Every minimal movement corresponds inside the learner-body to a distinguished neural state. Inside the learner-body we assume further a perception which represents the space as a 2-dimensional grid with highlighted values in case of occupied positions. Within such a grid-vision the different movements have a distinguished geometrical meaning.

3. Grounding of Perception: It is assumed that there is a grounding of subfields of the perceptual field correlated with timely correlated motor states. Let L be the movement of a body 'to the left' in the room, then we label the corresponding neural motor state ML. The corresponding perceptional unit is represented as PL and the grounding of PL into ML is given as $CL: PL \longrightarrow$ ML assuming that there is some cognitive state $CL$ available representing this connection.

4. Imitation Learning: We require from the learner that he is able to generate by itself minimal movements which can be concatenated into sequences called movements. The learner can distinguish between those movements caused by himself and those caused by another body. Furthermore he can compare sections of his perceptually represented movements with sections of movements from another body.

5. Self: It is assumed that the distinction between the other body and the learner body is based on the fact, that there is a perceivable correlation between perception and motor states represented as some cognitive state. Therefore all the motor grounded perceptions can be classified as learner body related or ME compared to the non motor grounded perceptions as other or YOU.

6. Storage, Movement Chunks: It is assumed that all cognitively represented movements can be stored as distinguishable units forming sequences of length 1 until $k$ called movement chunks. A context anchor is attached to each movement chunk to indicate the position from where these movements have been started (in idealized grids the context anchor is given by the coordinates of the grid-point assuming the y-direction as 'O := Oben := above := ahead. In more elaborated spaces there can be properties of the space around which function as a key to the starting position). Every such movement chunk has additionally a label representing the percentage of world similarity with movement sequences observed with other bodies. (Possible extensions can be other value labels indicating special fitness for special targets).

7. Dreaming: Periodically a dreaming operation will happen. A randomly selected subset of those movement chunks which are below a certain threshold $\epsilon$ of world similarity which will be manipulated according to a crossover or a mutation operation.

8. Abstraction: If $mc_{i}, mc_{j}$ are two different movement chunks then there can also exist a chunk connector $mcc(mc_{i}, mc_{j},\alpha)$ representing the sequence of both movement chunks with world similarity value $\alpha$. Chunk connectors can be recursive like $mcc((mcc(mc_{i}, mc_{j},\alpha),mc_{r}, \beta)$.

The assumptions labeled Storage, Movement Chunks as well as Dreaming can theoretically be understood as parts of a genetic programming paradigm (cf. []).