Evolving crawling!

Crawling

Game Fitness =  The difference from the start position and the final position of the centre of gravity in cartesian coordinates. 

Sensor Atom Parameters: Inputs sensors to CPG 

Parameters of SHC:  CPG parameters of CPG. 

Motor Atom Parameters:  Output motors of CPG.

A critical question is how to encode motion? How to encode a behavioural sequence that is going to lead to high variance in the game fitness? 

How to encode motion? 

" A traditional view of the motor cortex in the primate brain is that it contains a map of the body

arranged across the cortical surface. This traditional topographic scheme, however, does not capture

the actual pattern of overlaps, fractures, re-representations, and multiple areas separated by

fuzzy borders. Here, we suggest that the organization of the motor cortex, premotor cortex, supplementary

motor cortex, frontal eye field, and supplementary eye field can in principle be understood as

a best-fit rendering of the motor repertoire onto the two-dimensional cortical sheet in a manner that

optimizes local continuity." 

https://www.princeton.edu/~graziano/neuron_07.pdf

Well, it seems that regions of the motor cortex when stimulated result in immediate production of complex motor sequences, e.g. climbing/leaping, reaching to grasp, defence, manipulate in central space, etc., see `Figure 5. from the above paper below. 

How are the above complex sequences of action encoded?

Primary MC: Encodes force of muscle contraction and direction of movement of limb (Georgopolos) as a sum of a population of vectors. Extent and speed of movement. [http://openwetware.org/images/6/6a/2012MotorCortex.pdf]

Postures are encoded by pre central cortex.
[http://www.cell.com/neuron/abstract/S0896-6273(02)00698-0]

Premotor cortex seems to do conditions on action classifiers, i.e. they detect hidden-state in the environment, e.g. 'there is an edible peanut around' ('Motor Set' it is traditionally called).

Supplementary Motor Area: Programs complex movements???, Bilateral motion, transform kinematics to dynamics???

Association Cortex:

Read the following review! 
1. Overall motor control organization. 
http://www.cisek.org/pavel/Pubs/CisekKalaska2010.pdf

2.
http://students.washington.edu/zanos/literature/Cortical%20plasticity/Motor%20cortex%20plasticity/Sanes%20Donoghue%202000%20-%20Plasticity%20and%20primary%20motor%20cortex%20%5BR%5D.pdf

3. Cortical control of motor sequences
http://www.sciencedirect.com/science/article/pii/S095943880600033X#

Hacking event in Dutchland, lovely, definitely going to present Janky robots. Staying in a tent. NOW THAT IS HACKING AT ITS BEST. Thanks Anthony Liekens for inviting me. 
https://ohm2013.org/wiki/Village:OpenGarage

VERA'S IDEA: Motor Molecules for Complex Action Sequences 

Another great idea from Vera which exploits the power of the Fluid Action Grammar (FAG!). She suggests that instead of there being a single motor atom controlling behaviour that the behaviour is specified by a sensori-motor molecule. This sensori-motor molecule replicates as a unit and specifies the behavioural sequence. Lets imagine a motor molecule for a complex (potentially recursively specified) action sequence. It starts with a sensory molecule in the standard way as these sensory molecules are the initiators of actions on the basis of the environmental cues. The sensory molecule activates and passes sensory information to a motor-sequence atom. This motor sequence atom is distinct from the standard motor atom we've been considering before. It specifies a non-blocking setAngles call and then activates another motor-sequence atom which is activated but which only does its action after some specific delay which is specified in it. In this way, a network of motor-sequence atoms becomes activated. Motor-sequence (MS) atoms can also have sensory conditions, i.e. take in sensory input and produce motor actions only if it is within some range, or the motor output can be a function of the sensory input to that MS atom, thus allowing CLOSED-LOOP as well as OPEN-LOOP control. 

Recursive control is also possible, i.e. A to  B to A. A has its timer reset whenever it is activated. MS atoms have timers unlike standard M atoms. 

The atom at the beginning, the starting atom, which is the sensory atom is a bit different from the previous sensory atom. It writes the position of the agent in the 3D space to memory. This is not directly used by the molecule but the game to determine the fitness of the SM molecule. 

The game atom associated with this SM molecule is to maximise the output of this first sensory atom. 

This project will give us an opportunity to try the structural mutations, because functional changes depend here on structural mutations as well as specific atomic mutations. 

OK, main original molecules coded up. Sensory to MotorP to MotorP is the organisation.

TO DO: Program structural mutants of motorB atoms. 


OK, first attempt programmed below. 

https://github.com/ctf20/DarwinianNeurodynamics/commit/13b0503377d5579e110c341fb7ba98dadef4fa86

TO DO: 1. Increase population size. 2. Ensure a decent fitness function is being used to score molecules. 3. Get some visibly adaptive behaviour evolving!