notes

sbt
run

There are two world Planes - front and back

+ world needs background spatial markers for navigation


What is the API to the sensorimotor system? 

Models:
  blocks with friction, density, texture

   server can implementat 'sticky' blocks



Motor output:

System has two arms. Each has two joints and a hand:

  Gross motor action: upper arm, moves to a quadrant of screen, say 1/10 quantized

  Fine motor action: lower arm, moves 1/10 of quadrant

  Rotator Joints at the shoulder and elbow, i.e., there is redundancy for reaching a point but that
  allows for better reuse of relative motion schemas

  Gross and fine joints have control over force left,right, up/down

  "Lift" pulls a block out of the plane so it can be moved over other blocks.

  - The motor system provides both force and a targeted 'move to proprioceptive position' function. E.g., 
  we've got a fairly high level motor system.

  hand: grasp, ungrasp, rotate left, rotate right
    grasp and rotate force is controllable over some range
     -- the 'grasp' operation gives a force feedback of "grip" when there is an object touching, with force sensation
        proportional to the gripping force, like when you squeeze a block with you fingers.


    HEAD: rotatation , translation

    Gaze: 10x10 gross angle quantization
    	  10x10 fine angle position

Also see Vision system, there is 'motor control' of internal pipeline vision filters, you can
choose which ones go into pipeline

Vocal: phoneme output

Sensor inputs:

  Proprioception: 
     10x10 location for gross motor system - divided into x,y axis? 
     10x10 location for fine motor system 
     shoulder and elbow joint position

Force/touch

     force on shoulder joint, x, y, rotation

     force on elbox x,y,rotation

     force on wrist join rotation, x, y 

     force on grasp

     touch on fingers, texture, temperature, heat conductivity, on fingers. 

     

encoding of magnitudes; should we encode as V>n ? 

v= 6

n = >1 >2 >3 >4 >5 >6 >7 >8 >9 >10

    1  1  1  1  1  1  0  0  0  0




Vision

  modules:

  broad-area vision - blob detection per quadrant, a couple of bits of color/grayscale

  fovea operations:

     closed area fill - is area at fovea closed? 
     are there any bits on at the fovea ; color, density, texture (image morphology bank of filters)


     edge detection -- vertical edges, horizontal edges, diagonal edges (in 30 degree increments

  filter system (image morphology)
    prescale
    erode/dilate
        

   motion tracking; attention focus points on fovea and broad-area peripheral vision  when motion is sensed
     you can enable tracking as an action on the fovea; you then get the gaze moved around for you as it tracks an object
     that is visible.


     Motor action: focus eye on object center of mass. Send hand to gaze position? 

     +      Motion map: at each timestep, compute which objects are moving, light up sensors in a motion quadrant map.

     +      Motion at fovea - has more directional info, velocity. 

     When the head moves, mark all quadrants which have objects as in motion? Or is that something that
     we should handle in the vision system, and factor out by the time it reaches schema mechanism? What can you
     learn from moving head and seeing an object motion vector opposite direction? I see the value of understanding
     its final location, i.e., it's static position visual sensor, but what value is the momentary motion illusion?

     

   trajectory - when motion detected, a motion vector input is stimulated


   Foveation: vision system will pull focus of attention to nearest object, and give a proprioceptive
   sensation of moving the gaze by dx,dy.  Vision hardware can be asked to cycle to nearest objects, in one
   of four directions; left,right,top,bottom. You get sensation of how far the gaze moved to get to next object.
   That way you can trace out the location of several nearby objects, and get a sensation of how they are 
   relatively positioned.



Audio

  Basic frequency burst (pop, click, boom), localization from auditory system to a quadrant
  Phoneme inputs



================================================================

How is sensorimotor info passed to schema engine? 

class SensoryInput {
      int id; // a unique id for this input
      boolean value;
      boolean changed;
}

class MotorAction {
      int id; // unique id for this output
      boolean value;
}

class WorldState {

    HashMap<String,SensorInput> inputList = new HashMap<String,SensorInput>();
    
    HashMap<String,MotorAction> outputList = new HashMap<String,MotorAction>();


}







================
Questions:

How do infants learn to keep still? I.e., it's important to only be moving a few things at time, initially.
So you need to keep most of the body still. And for fine motor activity need to keep the gross motor joints still.

How much of that is built in? You need to set the opposing muscles at
exactly the same force to maintain constant proprioceptive input. In
terms of schema, what are the interesting schemas being learned there,
to maintain that invariant input? Why is it interesting to the Schema mechanism to learn
to keep limbs still by default?  Or to keep head aligned with an object to keep it centered in fovea? 


call David McDonald about language


================================================================
export CLASSPATH=lib/controlP5.jar:lib/core.jar:lib/gluegen-rt.jar:lib/jbox2d_2.1.2_ds_v2.jar:lib/jogl-all.jar:lib/jruby-complete-1.7.4.jar:lib/pbox2d.jar:lib/trove4j-3.0.3.jar:lib/typesafe-config.jar:target/scala-2.10/jschema_2.10-1.1.jar:lib/log4j.jar

jirb 




Web API

http://127.0.0.1:8080/
http://127.0.0.1:8080/items/map
http://127.0.0.1:8080/raw/map




================================================================
Main loop - select an action, perform it once, wait n cycles to
attribute effects to it.

Need to remember last action activated, but only activate it once.

+ Need to bias the gaze to stay centered where the head is.

+ add grasp reflex: = when hand touched, grasp it
 -- preinitialize schema for this ?

+ Do some whole areas of the brain go idle when doing learning, so you might be biased towards just learning
associations between motion and touch, or touch and vision? So that not too many are active at once
to swamp the learning mechanism? 

+ Embellishment - when spinning off new (result) schema B from parent A, add B's synthetic item to the item ignore list
of A. Because any child schema's synthetic item will always be correlated to parent activation.

>>>>Check if Drescher already covered this case with suppress/override heuristic ...

    Hi Henry. Yes, I think that should be suppressed. Also, no synthetic
    item should be defined for a schema with an empty result (a synthetic
    item reifies the validity conditions of a schema, and a schema with an
    empty result has no validity conditions).

    For a schema with a nonempty result, it shouldn't be possible that the
    schema's own synthetic item is discovered to be a result of that
    schema's activation. At most, that schema's activation should
    transition the schema's synthetic item from Unknown to On, not from
    Off to On; and the former transition does not count for the
    extended-result correlation statistics. (If activating a schema were
    to turn its own synthetic item On--which is to say, if activating the
    schema were to achieve the schema's validity conditions--then the
    schema would always be valid, in which case its own synthetic item
    should always be On, rather than being designated as a result of the
    schema.)

+ Check what are the exact conditions under which a schema's synthetic item goes to what state? 

+ "My implementation used an ad hoc method that was tied to its
space-limited statistics collection method. But the real way to do it
is to use a threshold of statistical significance. So just pre-compute
a lookup table that says what the minimum correlation is that can be
supported by a given sample size.
"

+ I'm getting positive feedback in selection/creation; I pick a schema at random to execute,
but if they are based on HAND_UP, then they spin off some HAND_UP schemas, so the random selection
is biased towards moving hand up.

+ How does a touch-grasp  reflex work? Is there a schema for it? it operates outside of schema mechanism,
as an instructor? 

+ How do I implement this (4.1.2) "These statistics are tabulated over
a number of trials in which the action is tak- en, and a number of
trials in which it is not; the more trials there have been, and the
more discrepancy there is between the two probabilities, the sooner
the machinery will detect the difference (see section 5.2.2). The
sampling is weighted toward the most recent trials."


+ When a parent spins off a child result schema for an item, do you reset that item's counters to zero in the parent schema 
extended result? 

sudo apt-get install xvfb
Xvfb :2 -screen 0 1024x768x24 &
export DISPLAY=localhost:2.0
java -jar jschema-assembly-1.1.jar



+ make items record timestamp of last transitions. in realtime world we need this to span intervals of time
in the simulation. 

+ motor actions - should be "turn on" "turn off" so you can "sweep arm right" which will continue until
you "stop arm. 

How should these be represented/implemented? Do we have unit motion "move left one" as an atomic action, 
or do we have "start moving left" and "stop moving" ? 
If you have 'start moving' and 'stop moving', how do you learn to move small unit amounts? 


++ When gaze moves, need to damp out visual deltas (item transitions) for a frame?
Which items need to damp if any?


+ Q: When a schema predicts a result item, is it predicting that the item will be on (or off) or
is it predicting that the item will transition from off->on? 
Is there a different kind of schema for predicting a transition vs. predicting what the final state of the
item is? 
A: A schema which says  ~LIGHT-ON/ FLIP_SWITCH/LIGHT-ON predicts a transition of the value

+ refactor marginal attribution to loop over all items in outer loop, 
and only update on changing items.

+ sort SensorItems by transition time, then the copysms loop only has to look back as far as the last action time.
Use TreeSet

java_import "com.beartronics.jschema.WorldState"
w = WorldState.new
i = w.inputsByTransitionTime

w.setClock(20)
s1 = w.setSensorInput("foo", 0, false)
w.setSensorInput("foo", 0, true)
w.setClock(30)
s2 = w.setSensorInput("bar", 1, false)
w.setSensorInput("bar", 1, true)
w.setClock(40)
s3 = w.setSensorInput("baz", 2, false)
w.setSensorInput("baz", 2, true)
w.setClock(50)
w.setSensorInput("foo", 0, false)





+ Action cycle should be that an action is taken, then system waits for input transitions to
settle down, with some max time to wait for things to quiesce, then does learning step.



+ Should we restrict spinoffs from schemas tht have no context items? 

Or do schemas with no context (only results) not get synthetic items? Because they're not
potentially useful enough for that? 

+ do we need a pool of schemas, from which new ones are pulled, and when full, old ones
are recycled? 


+ need to get action selection to repeat recently successful spun off schemas? 

Gson gson = new GsonBuilder()
.excludeFieldsWithModifier(Modifier.STATIC)	
.enableComplexMapKeySerialization()
.serializeNulls()
.setFieldNamingPolicy(FieldNamingPolicy.UPPER_CAMEL_CASE)
.setPrettyPrinting()
.setVersion(1.0)
.create();


gson = GsonBuilder.new().excludeFieldsWithModifier(Modifier.STATIC) .enableComplexMapKeySerialization() .serializeNulls() .setFieldNamingPolicy(FieldNamingPolicy.UPPER_CAMEL_CASE) .setPrettyPrinting() .setVersion(1.0) .create()

gson = GsonBuilder.new().excludeFieldsWithModifier(Modifier.STATIC).create()


import java.lang.reflect.Modifier;

Gson gson = new GsonBuilder()
    .excludeFieldsWithModifier(Modifier.STATIC)
    .create();

http://stackoverflow.com/questions/4231092/gson-illegalstateexception

use Trove HashMap (hashset?) implementation

================

implement context spinoff

if you have 

p/a/x

and there is an xy/b/z someplace , and the marginal attribution for p/a/x 
notices that

p/a/x  { xy is correlated }

do you then spin off 

p/a/xy ? 

+ What if:

  you have p/a/x, and its extended result sees that the "ghi" of some other schema ghi/b/z is correlated?

  Does that spin off p/a/ghi ? Is there some info we're losing about the result x? 


Does a 'transition' of a conjunct 'xy' or 'ghi' mean that they all transition from off to at the same time ? 



+ Should the simpler schemas (less context and result items) have a higher number of trials threshold before
they spin off, on the assumption that they are noisier? 

================================================================


Need to get adjacency map case working -- start with 1-d hand motion.

Is there any advantage to the approach of keeping proprioceptive X and Y separate, vs a single hand@x,y sensory input? 


- make grossx, finex be zero based, no negative values, too confusing to read

================================================================

When you spin off a new schema (in particular a new schema with a
context item), what if any parts of the extended context and result do
you copy to the new child schema, and what entries in the extended
context and extended result if any do you reset in the parent schema?


================
Make a JSON config file for all params so they are in one place!

================

Seems to be bug with move hand logging effect of hand position two units away

Add debug logging to extended-result ,to show what input delta we're
tallying. Check if we're getting correct results 

- then move to extended result, log what preconditions we're tallying. 

Q: When you spin off a new schema with a new context item, do you clear the parent schema's extended result?

Q: When you spin off a new schema with a result item, do you clear just that result counter from the parent's
extended result, or do you clear all extended results?

#####
Is the criteria for "succeeded" if the result is found in the predicted state, or it *transitions* to the predicted state?
I.e., is the result for a schema a prediction of value (regardless of whether the value was already in that state) or of a transition? 



================================================================
	p. 76
    Suppose, in the example just discussed, that
    the context-relevance of d to schema /a/x is not obscured; the schema' s extended
    context discovers this relevance, leading to the construction of the schema d/a/x.
    The extended-context slot for d in /a/x records that a schema has been spun off
    from that schema for that (positively included) item. The following embellishment
    then occurs:

    • All correlation data in all extended context slots of the schema /a/x are reset
    to zero.

    • Subsequently, whenever /a/x is activated and d is On, the updating of all extended
    context data for that trial of /a/x is suppressed. The effect of this embellishment
    is that the extended context of /a/x now maintains correlation
    data only for trials for which d is not On (resetting the data erases correlations
    that had been tabulated without this condition). Thus, when d is on,
    attribution is deferred from /a/x to the more specific applicable schema d/a/x.
     That schema, of course, can update its own extended context data for the
    trial, leading to the eventual construction of def/a/x.

+ does result spin off reset all counters?? 
  how do we inhibit redundant result spinoff? 

 -- talk to Gary 

+ TODO p 85 child schema reports its applicability to the parent, which can turn on it's synthetic item.
also if an ovveride conditions obtains in the child, it can turn off the parent's synthetic item



================================================================
+ When we build conjunct items (for use as targets for result) do we need to make sure not to use them
as context items for other schemas? Because a conjunct item would be redundant with the schema context list that created it?

+ be careful to update the conjunct item values exactly:  lastPosTransition, lastNegTransition, value, prevValue
all need to be set properly for taking of both context and result statistics.

+ Build conjunct items instead of conjunct lists  -- easier to deal with

+ use X,Y proprioceptive items, and visual items, instead of discrete separate x and y items


================
When spinoff happens, log everything, put in string in schema for easy access, se we see counter values

-- look up override conditions, how are those represented in parent?
- remember override conditons when doing planning



================================================================


(a + b)! (c + d)! (a+c)! (b+d)!
_________________________________
       a! b! c! d! n!


             no-action  action

transition     a          b

no-transition  c          d  
18 Item-18 #18:hand1.gross.(-2,-1)   ^ NaN [A: 2.00/192, !A: 7.00/5268],  v NaN [A: 3.00/192, !A: 6.00/5268]

a positiveTransitionsNA = 7
b positiveTransitionsA = 2
c numTrialsActionNotTaken 5268
d numTrialsActionTaken, = 192


   -3 -2 -1  0 

-3  .  O  .  .

-2  x  .  .  .

-1  .  .  .  .

~[]/[Action-1 HAND1_RIGHT]/Item-14 PRIMITIVE #14:hand1.gross.(-2,-3):~null]

sense=false:= 2>Item-2 #2:hand1.gross.(-3,-2)  </a> On 0.000000 [Succ.: <b>0</b>, Fail: <b>5</b>],  Off 0.285714 [Succ.: <b>2</b>, Fail: <b>7</b>] 1/0 0.000000 0/1 Infinity <b></b> <b></b>






saveStage() and loadStage() methods


================================================================
1.9.3-p484 :002 > Exception in thread "Animation Thread" java.lang.Error: spinning off OFF context item [Schema 13 []:~[]/[Action-1 HAND1_RIGHT]/null:~Item-10 PRIMITIVE #10:hand1.gross.(0,1)]


	   	  	       	      ITEM was	ON/suc	ON/fail 	  OFF/suc      OFF/fail	   
0 Item-0 #0:hand1.gross.(-1,-1)	      On NaN	0	0	Off 0.90	18	2	 1/0 NaN	 0/1 NaN
1 Item-1 #1:hand2.gross.(-1,-1)	      On NaN	0	0	Off 0.90	18	2	 1/0 NaN	 0/1 NaN
2 Item-2 #2:hand1.gross.(-1,0)	      On 1.00	1	0	Off 0.89	17	2	 1/0 1.117647	 0/1 0.894737
3 Item-3 #3:hand2.gross.(-1,0)	      On NaN	0	0	Off 0.90	18	2	 1/0 NaN	 0/1 NaN
4 Item-4 #4:hand1.gross.(-1,1)	      On 0.00	0	2	Off 1.00	18	0	 1/0 0.000000	 0/1 Infinity
5 Item-5 #5:hand2.gross.(-1,1)	      On NaN	0	0	Off 0.89	17	2	 1/0 NaN	 0/1 NaN
6 Item-6 #6:hand1.gross.(0,-1)	      On 1.00	1	0	Off 0.89	16	2	 1/0 1.125000	 0/1 0.888889
7 Item-7 #7:hand2.gross.(0,-1)	      On NaN	0	0	Off 0.89	17	2	 1/0 NaN	 0/1 NaN
8 Item-8 #8:hand1.gross.(0,0)	      On 1.00	1	0	Off 0.89	16	2	 1/0 1.125000	 0/1 0.888889
9 Item-9 #9:hand2.gross.(0,0)	      On NaN	0	0	Off 0.89	17	2	 1/0 NaN	 0/1 NaN
10 Item-10 #10:hand1.gross.(0,1)      On 1.00	5	0	Off 0.86	12	2	 1/0 1.166667	 0/1 0.857143
11 Item-11 #11:hand2.gross.(0,1)      On NaN	0	0	Off 0.89	17	2	 1/0 NaN	 0/1 NaN
12 Item-12 #12:hand1.gross.(1,-1)     On 1.00	2	0	Off 0.88	15	2	 1/0 1.133333	 0/1 0.882353
13 Item-13 #13:hand2.gross.(1,-1)     On NaN	0	0	Off 0.89	17	2	 1/0 NaN	 0/1 NaN
14 Item-14 #14:hand1.gross.(1,0) TRUE On 1.00	3	0	Off 0.88	14	2	 1/0 1.142857	 0/1 0.875000
15 Item-15 #15:hand2.gross.(1,0)      On NaN	0	0	Off 0.89	17	2	 1/0 NaN	 0/1 NaN
16 Item-16 #16:hand1.gross.(1,1)      On 1.00	4	0	Off 0.87	13	2	 1/0 1.153846	 0/1 0.866667
17 Item-17 #17:hand2.gross.(1,1)      On NaN	0	0	Off 0.89	17	2	 1/0 NaN	 0/1 NaN

This schema predicts that when we move HAND_RIGHT, Hand@(0,1) will be OFF.
This schema of course succeeds almost all the time. 
Why did system propose it? 
Should the criteria for "schema succeeds" be that result item is in predicted state, or that it does a transition to predicted state?



================================================================
 updateContextItems Schema [Schema 14 []:~[]/[Action-2 HAND1_UP]/null:~Item-4 PRIMITIVE #4:hand1.gross.(-1,1)] succeeded=true
 item Item-4 PRIMITIVE #4:hand1.gross.(-1,1), onValueReliability: 1.0, offValueReliability: 1.0, on_succeeded: 1, on_failed: 0, off_succeeded: 19, off_failed: 0
 handleActivation [Schema 14 []:~[]/[Action-2 HAND1_UP]/null:~Item-4 PRIMITIVE #4:hand1.gross.(-1,1)] applicable=true, succeeded=true
 Activated schema: [Schema 12 []:~[]/[Action-12 GAZE_DOWN]/null:~null]


I'm starting with the bare schemas of primitive actions HAND_UP, HAND_DOWN, HAND_LEFT, HAND_RIGHT, as well as some other actions which don't affect anything, like GAZE_RIGHT, NULL_ACTION, HAND_GRASP. 

I've set the sensorimotor system so that the hand moves just to nine discrete positions where  -1 < x < 1, and -1 < y < 1, so it perceives boolean hand location sensory-items like HAND@(0,0) HAND@(-1,1), HAND@(1,0) etc.


So for example if you have HAND@(0,0) = true, then all other HAND@(x,y) are false, and if you move the hand with HAND_RIGHT, you get a negative transition on HAND@(0,0), and positive transition on HAND@(1,0). (note, if the hand is at 1,1, and you move it right, it can't move any further, so there is no change in location, and HAND@(1,1) remains on)

After running for a while and taking extended-result statistics on the bare schemas, I see a spinoff schema get created:

/HAND_UP/~HAND@(-1,1)

Which says if I move HAND_UP, then sensory item HAND@(-1,1) turns off, designated as ~HAND@(-1,1). 

~HAND@(-1,1) could happen in the cases where HAND@(-1,1) and the any of the actions HAND_UP, HAND_DOWN, or HAND_RIGHT are taken.
HAND_LEFT would not cause any change in this case because the hand is at the left extent of it's travel, as we are only letting the hand x and y positions range from -1 to 1 in this experiment.

Once the schema /HAND_UP/~HAND@(-1,1) has been created, then I am getting a problem with the extended context creation.

Because in lots of cases, when this schema executes, it is regarded as having succeeded, because in almost any hand location,
if you move the hand in any direction, you will find that HAND@(-1,1) is false. Because the criteria for 'success' I am using 
now for a schema is that after the action is taken, the predicted state of the items in the result set are satisfied. So I'm
not looking for the result items to make transitions to the predicted values, just that they have in those predicted values after the action is taken.

The current criteria the extended result mechanism is using to define 'success' is threefold:

1) The schema's context is satisfied (always trivially so in this case, because the context is empty)

2) The action is taken

3) The result conditions are satisfied after the action is taken

So in this case I start getting lots of spinoff schemas from the extended-context phase of the form

~HAND@(x,y) / HAND_UP / ~HAND@(-1,1)

I.e., for almost every sensory item in the extended context of HAND_UP/~HAND@(-1,1), that item will be
OFF when after schemas executes, and the schema pretty much always succeeds, since HAND@(-1,1) is usually false, so we increment the OFF_WHEN_SCHEMA_SUCCEEDS counter
for those items, and soon the marginal attribution mechanism decides, for most x,y locations, that ~HAND@(x,y) causes this schema to reliably succeed.

So I'm getting a bunch of spurious schemas with context items of the form

~HAND@(0,0) /  HAND_UP / ~HAND@(-1,1)
~HAND@(1,0) /  HAND_UP / ~HAND@(-1,1)
~HAND@(1,1) /  HAND_UP / ~HAND@(-1,1)
...

Should I be doing some other criteria for success, like actually
checking that the schema's result items actually make a transition to
predicted values in that timestep, rather than just checking their
final states match the schema's result set predicted value?





================================================================


when spinoff context child schema, save the extended context of parent for debugging

- do we have a sign error in spinoff of context OFF items? 



__________________________
|       |        |        |
|       |        |        |
|(-1,-1)| (0,-1) | (1,-1) |
|       |        |        |
|       |        |        |
|-------+--------+--------|
|       |        |        |
|       |        |        |
|(-1,0) | (0, 0) | (1,0)  |
|       |        |        |
|       |        |        |
|-------+--------+--------|
|       |        |        |
|       |        |        |
|(-1, 1)| (0, 1) | (1, 1) |
|       |        |        |
|       |        |        |
|_______|________|________|



================
The system comes up with

/HAND_RIGHT/~HAND@(-1,1)

At that point, every single hand1 position will satisfy that schema, so you get
one of each

HAND@(x,y)/HAND_RIGHT/~HAND@(-1,1)

These schemas don't hurt anything I guess, but it seems like they aren't very useful. 
They are good for turning off HAND@(-1,1) I guess. But it seems like there will be a lot of them;
For all (x,y) , there will be x*y schemas which say how to end up with an  OFF state for hand@(x,y).



hand1.gross.(0,0)/[Action-0 HAND1_LEFT]/~hand1.gross.(0,1)




__________________________
|       |        |        |
|       |right->!|        |
|(-1,-1)| (0,-1) | (1,-1) |
|       |        |        |
|       |        |        |
|-------+--------+--------|
|       |        |        |
|       |        |        |
|(-1,0) | (0, 0) | (1,0)  |
|       |        |        |
|       |        |        |
|-------+--------+--------|
|       |        |        |
|       |        |        |
|(-1, 1)| (0, 1) | (1, 1) |
|       |        |        |
|       |        |        |
|_______|________|________|



___________________________
|       |        |        |
|       |left->! |        |
|(-1,-1)| (0,-1) | (1,-1) |
|       |        |        |
|       |        |        |
|-------+--------+--------|
|       |        |        |
|       |        |        |
|(-1,0) | (0, 0) | (1,0)  |
|       |        |        |
|       |        |        |
|-------+--------+--------|
|       |        |        |
|       |        |        |
|(-1, 1)| (0, 1) | (1, 1) |
|       |        |        |
|       |        |        |
|_______|________|________|

================================================================

Section 6.4 Beginnings of the persistent object concept

Important idea: Schema serves as a "host" to a synthetic item


Imitation: very important for learning by example

 4.3.2 Implicit activation and the representation of external actions
 As noted in section 4. 1 .2, the marginal attribution facility considers a schema to
 have been implicitly activated if the schema's action is initiated when the schema
 is applicable, even if that schema was not selected for activation, and thus was not
 responsible for the action 's initiation. Composite actions carry implicit activation
 one step further. A composite action is considered to have been implicitly taken
 whenever its goal state becomes satisfied—that is, makes a transition from Off
 to On—even if that composite action was never initiated by an activated schema—
 in fact, even if the goal state's achievement is due to external events entirely
 uninfluenced by the mechanism. Consequently, a schema whose action is composite
 is implicitly activated each time its action's goal state becomes satisfied
 when the schema is applicable. Marginal attribution can thereby detect results
 caused by the goal state, even if the goal state obtains due to external events.
 Designating external events as actions combines with activation hysteresis
 (section 3.4.2) to promote imitation by the schema mechanism of external events
 that correspond to extant schemas. Hysteresis promotes the activation of a schema
 that has been activated recently. Hysteresis applies even to implicitly activated
 schemas, so if a schema is implicitly activated because an external event
 achieved its action's goal state, the schema's chances for selection for explicit activation
 are thereby boosted; its explicit activation would then repeat the achievement
 of that goal state.

+ How do you come up with the abstraction "sideways"? 
As in "move it sideways to make free space" or "move it sideways to get it closer to X".
 You can push or pull it to get the same result.