Game ideas + Bilateral copying of functions in the brain?

Choose 10 games that Nao should play, encode them, and evolve solutions to show it can be done with hand designed (non-evolved) games, using the solution representations and variation operators we're using. This is the critical first step prior to running the co-evolution itself. 

Game 0: Maximise the absolute sum of an appropriate accelerometer reading over a trial.

Game 0a, as above for a different axis. 

Game 0b, as above for a different axis.

Game 1: Roll-over, i.e. maximise the change in the appropriate accelerometer axis in a fixed time

Game 1a: Roll-over, i.e. maximise the change in the appropriate accelerometer axis in a fixed time 

Game 1b: Roll-over, i.e. maximise the change in the appropriate accelerometer axis in a fixed time 

Game 3: Maximize the gyroscope value in the same axis as the accelerometer above.
Game 3a, as above for a different axis.
Game 3b, as above for a different axis.

Game 4: If the sensory vector is novel, i.e. not close to other sensory vectors in the already experienced sensory space, then make it a goal of a new game, with the fitness being the inverse of the Euclidean distance from that state summed over the trial. This is a Lamarckian game variation operator because the new game genotype is informed by the environment. So actually this generative process specifies a on-line mechanism for the generation of new games during experience. This is related to SAGG-RIACs search in sensory goal space and Goron's Hierarchical Curiosity Loops. It is a process of exploration in static sensory goal space.

Game 5: A game is specified by a Fourier decomposition over a sensory vector. Fitness is distance from that desired position at the specific time in the trial, e.g. game may be to make a sinusoidal movement of the elbow with a particular period, or to make a synchronised sinusoidal movement of both elbows at a certain period, in or out of phase, or a square wave movement in the same period/phase relationship. Controllers that would suit this would e.g. be LSM atoms.

Game 6: Try to maximise the amount of some particular colour in the visual field of one of the cameras. The more this is maximised throughout the trial, the more effective the motor molecule responsible. This may of-course, benefit motor molecules with access to vision, but not necessarily if some environmental element has a strong correlation between colour and some other modality that is detectable by a motor molecule without vision.

Game [Molecule] 7: Punish motor molecules that require a large joint stiffness to carry out their tasks. This is a kind of game modifier atom really isn't it. It modifies a game by taking a set of sensory (joint stiffness inputs) and punishing the sum of stiffness over the trial. This is a kind of sub-function of the overall game molecule. A game molecule is the combination of the standard game atoms above, and this game modifier atom. Game molecules consist of joined game atoms. For now, these are LINEAR combinations of game atoms, but more complex molecules are conceivable.

ONE GAME ATOM FUNCTIONS HAVE BEEN EVOLVED, GAME MOLECULES CAN BE FORMED WHICH COMPRISE COMBINATIONS OF GAME ATOMS. FOR NOW WE RESTRICT ATTENTION TO LINEAR COMBINATIONS OF GAME ATOMS. 

Wonderful question arises:

If you learn a skill on the right hand, does it increase the rate of learning it in the left hand? For which kinds of motor skill is the following true? 

1. Learning act on L hand  means immediately you can do it on the R when you try. IMPLIES shared representation does both. 

2. Learning act on L hand improves rate of learning of act on R. hand IMPLIES copying. 

3. Learning act on L hand does NOT increase rate of learning on R. IMPLIES relearning from scratch, no information transfer, no copying. 

What about for learning variants of the act on the R compared to that you learned on the L? Which set of mutants can be learned faster on the R than on the L? Hypothesis: The set of variants that learning on the L helps learning on the R for is the exploration distribution available to the copying operation from L to R brain representations.

e.g. Learning to play violin with L hand, how much does it help to re-learn on R. Of-course, this is too complex, as there are shared non-bilateral skills, i.e. reading music, that help both L and R side learning. 

Chrisantha Fernando: Problem is as [i think] Graham White pointed out, supporting Geraint's position, that there is a 4th possibility that that X->L => R<-x->L is distinct from the operation X->L => X->L + X'->R. This is formation of a new functional association rather than copying and I agree, I think it causes problems for the experiments above. The thinking cap is now about how to distinguish the above two processes.

Nathaniel Virgo: How about this to distinguish the two possibilities? I don't think it works very well (for reasons I'll explain) but maybe it's somewhere to start from. 

First, get people to learn a suitable task with one hand (say right). Then get them to learn it with the other (say left). So far so good. But then you get them to learn a new task with the *left* hand. This task should be similar enough to the first task that they learn it more rapidly than they would if they hadn't learnt the first task. (Maybe drawing two different kanjis would be good for the two tasks.) Then you get them to re-learn the second task with the right hand.

Then take a second control group and just teach them the first and then the second tasks using the right hand, but nothing using the left. What we want to compare is the speed with which the two groups learn the second task with the right hand.

If the architecture is L <- -="" x=""> R then teaching the first task with both hands builds connections between L and X, and X and R. Then teaching the second task with the left hand effectively just teaches X a new skill, so there's little if any work to do in order to transfer it to the right hand, since the connections are already there. So under this hypothesis I would expect learning the second task with the right hand to be much quicker in the test group than the control group.

On the other hand, if the architecture is X->R; X'->L then when you teach the second task with the left hand, you're only teaching it to X'. Then when you teach it again with the right hand, you have to copy the new information in X' back to X. Therefore in this case I would expect much more of a delay in learning the second task with the right hand.

Having written all that, it occurs to me that the initial copying of X to X' could build up some kind of "copying connections" between X and X', which would make subsequent copying between the two much more rapid. Then learning the second task with the right hand would still be much faster in the test group than the control group. This kind of possibility might make it very difficult to distinguish these types of hypothesis experimentally.

CF: Thanks Nathaniel. Brilliant idea. Another issue: 1. There may be some work transferring the new skull from X to the R effectively. In fact, they new skill may be conceptualised as a connectivity pattern Y which may or may not partially overlap with X. So there may also be a delay from the association model not only the copying model. But this argument is less if the 'interrogation' paths (as Graham puts it) from X to L and X to R are least modified. 2. I agree with your final paragraph too. So really I think we are left with this! Copying can only be distinguished from association by the capacity for DIVERGENCE WITHOUT INTERFERENCE of the two copies.

Nathaniel Virgo Ah yes, I think interference could be key. If you can find two tasks such that learning the second reliably interferes with the ability to do the first, all you need to do is teach people the first task with both hands, and then teach the second in just one. If there's no interference it would definitely support the copying hypothesis.

Chrisantha Fernando: Yes, so the second task T2 interferes with T1 in the L hand. So if you teach both hands T1, and when one hand learns T2 it interferes with that hands T1 but NOT the other hands T1, then this argues that there are bilateral representations. BUT wouldn't that also support the hypothesis that there are specific weights between X and L and X and R with X remaining uncopied? There must be some conceptual error happening somewhere here, it goes around in circles. The weights themselves are informational entities, thats the thing, so Wleft and Wright could store the competence, and they would have formed independently without communication. T2 interferes with Wleft, when given only to the left, but produces no influence on Wright. So, the interference pattern you describe could also be explained by the associationist model couldnt it?




See...


https://github.com/ctf20/DarwinianNeurodynamics/commit/d451061141dcc251d3899181579e708ebe34f065


for LivingMachines2013_14 a version with saving and loading of genomes, where I'm about to implement the liquid state machine atoms in brian (a neuronal network simulator for python that seems very easy to use).