The problem of learning:

1
FSKBANN - Finite State Knowledge
Based Artificial Neural Networks
The problem of learning - the traditional
BP learning - independent (In, Out) pairs, -
procedural learning frequently (In, Out) pairs
are contect dependent, i.e. their
classification depends on examples learned
earlier.   Frequent problem is to learn a task
action sequence. Earlier tools must be extended
to permit also to refine a 'procedural domain
theory'.   In case of procedural domain theory
two problems must be solved - formulation
of the procedural domain theory, - network
representation of the procedural domain theory.
2
FSKBANN - Finite State Knowledge
Based Artificial Neural Networks  
Learning procedures - generally by accepting
advices. Special features of a domain theory
composed from advices and instructions is that
it is usually - fragmentary, - is
focused on a particular task, - can advice,
even prohibits, single or multiple steps.
3
For the refining of the procedural domain theory
two approaches will be used   FSKBANN
refining domain theory formulated as a
finite automaton.   RATLE refining advices
formulated in a procedural language, together
with the reinforcement learning.
4
FSKBANN
What will be a suitable tool ? Due to the
context dependency a network with memory (state
memory) is required, consequently a recurrent
network will used. SRN - Simple
Recurrent Network
5
FSKBANN
State (context) units Network remembers its
earlier state. Recurrent links are not present
in the learning. Activating an SRN network -
activating state (context) units (copying) -
activating other IN units accordingly to the
applied examples - activating other hidden
and OUT units by feed-forward propagation.
6
FSKBANN
To obtain a network able to refine finite
automata one must solve - the transformation
from the finite automaton into a suitable rule
set, and - the transformation of the obtained
rule set into a network.   Finite automaton ?
rule set transformation   State transition of
finite automaton Rule k A ??? B
Bk - Ak-1 ? k k - k1 ? k2 ....
7
Example Define rules suitable to implement binary
adder as a finite automaton. The automaton has
two states (C0) and (C1), and its inputs
are (In1, In2).   E.g. 0 0 1 1 In1 1 0 0
1 In2 0 1 1 0 Carry
------------------------------- 1 1 0 0
Out
Carry
In2
In1
?
Out
8
The state transitions are as follows  
(1,1) (x,0) (C0) ??? (C1) (C0) ???
(C0) (0,x) (0,0) (C0) ???
(C0) (C1) ??? (C0) (x,1)
(1,x) (C1) ??? (C1) (C1) ??? (C1)
9
Start
All-In-Yes
One-In-Yes
One-In-Yes
C0
C1
All-In-Yes
Nulla-In-Yes
Nulla-In-Yes
10
State transition rules C0 k ? C0 k-1 ?
Nulla-In-Yes C0 k ? C0 k-1 ? One-In-Yes
C0 k ? C1 k-1 ? Nulla-In-Yes C1 k ?
C1 k-1 ? One-In-Yes C1 k ? C1 k-1 ?
All-In-Yes C1 k ? C0 k-1 ? All-In-Yes
  Condition rules All-In-Yes ? In1-Yes ?
In2-Yes One-In-Yes ? In1-Yes ? ? In2-Yes
One-In-Yes ? ? In1-Yes ? In2-Yes
Nulla-In-Yes ? ? In1-Yes ? ? In2-Yes
Ki-Yes ? All-In-Yes ? C1 k-1 Ki-Yes ?
One-In-Yes ? C0 k-1 Ki-Yes ? Nulla-In-Yes
? C1 k-1
11
Translating rules into a network - as in case
of KBANN - implementing In and Out units, -
implementing hidden units and links acc. to the
rules, - implementing state units (2 units for
each state, 1 unit for the state in the current
time-point, and one for the next
time-point), - establishing time-delay recurrent
link, - listing state initial values what is
active at the beginning ?
12
(No Transcript)
13
Experiments Protein secondary-structure
prediction   Improving Chou - Fasman
algorithm (until now the best, here it will play
the role of the domain theory).   Protein 20
different aminoacid, aminoacid chain (gt 100) -
this is primer-structure, protein structure
3D (tertiary) structure, folding'   -
predicting tertiary structure - expensive,
time consuming - predicting secondary structure
predicting local environment of every aminoacid
- a-helix (a) - b-sheet (b) -
'random coil' region (-)
14
Know prediction (not learning) algoritms classify
with Accuracy of 50-58, however the result is
misleading. This app. the ratio of 'random
coil' in the chain, i.e. a trivial algorithm
every structure random coil' will classify
with app. the same accuracy.   Learning
approach - Feed Forward ANN, app. 63 of
accuracy, - hybrid solutions 62-69
15
Chou - Fasman algorithm (1) looks for
aminoacids, which, with high probability, are
parts of a-helix or b-sheet, (2) extends the
prediction to the neighbouring aminoacids
- part of a or b, based on - in itself, -
neighbours, default decision -
random coil
16
Chou - Fasman algorithm as finite automaton  
notation 'X_at_N' true, if the X aminoacid
is in Nth position from the actually predicted
place. Some rule example Nucleoid rules
init_helix ? ((Spos0..5 helix_former (AS_at_pos)) gt
4) ? (Spos0..5
helix_breaker (AS_at_pos)) lt 2 ) )
...... where helix_former or helix_breaker are
special combinartions of aminoacids.
17
Chou - Fasman algorithm as finite
automaton   Rules determining the helix
structure helix_break_at_0 ? N_at_0 ? Y_at_0 ? P_at_0 ?
G_at_0 helix_indiff_at_1 ? K_at_1 ? I_at_1 ? D_at_1 ? T_at_1 ?
S_at_1 ? R_at_1
? C_at_1 break_helix ? helix_break_at_0 ?
helix_indiff_at_1 etc. Rules determining the
sheet structure sheet_break_at_0 ? K_at_0 ? S_at_0 ?
H_at_0 ? N_at_0 ? P_at_0 ? E_at_0 sheet_indiff_at_1
? A_at_1 ? R_at_1 ? G_at_1 ? D_at_1 break_sheet ?
sheet_break_at_0 ? sheet_break_at_1   etc. Continuity
rules cont_helix ? ØP_at_0 Ù Ø break_hellix
cont_sheet ? ØP_at_0 Ù ØE_at_0 Ù Øbreak_sheet etc.
18
Finite automaton rules, based upon Chou-Fasman
automaton
19
helix k ? sheet k-1 ? init_helix helix k ?
coil k-1 ? init_helix helix k ? helix k-1 ?
cont_helix sheet k ? helix k-1 ? init_sheet
sheet k ? coil k-1 ? init_sheet sheet
k ? sheet k-1 ? cont_sheet coil k ? helix
k-1 ? break_helix coil k ? sheet k-1 ?
break_sheet coil k ? coil k-1 ? any
  Experiment 128 segment, S 21,600 aminoacid,
54,5 coil, 25,2 a-helix, 20,3 b-sheet
20
(No Transcript)
21
Comparing the solution - comparing ANN with
FSKBANN (C-F algorithm 57, ANN 61,5,
FSKBANN 63 based upon 80 protein example) -
FSKBANN better, whatever the training set, -
with more training proteins the accuracy will
surely grow. FSKBANN truly better, because it
predicts better the b-sheet structures.