Application of a "Bias" in AI Programming

[Dimster]

In an AI program, the decision making process often calls for the program to interpret an event and place a weight value to the results. As an example the program determines if the event increases or decreases a particular value. The increase/decrease could be used as a "weight" should that particular event be encountered in the future. I'm now exploring how to place a Bias to that weight however I'm not sure if the bias which I'm reading about in AI programming is one generated by the programmer or does the program itself determine the bias. Would a bias mean that I would need to consider the likelihood of the event occurring as I have a better idea of where I want the program to go or would a "normal" AI program explore the event more (ie how many times in the past or how relevant/impactful the event was) in order to measure a bias value?

[FellippeHeitor]

Deep question. Hope someone will have a good reply for you soon (I have no idea).

[TerryRitchie]

I can barely comprehend my own intelligence (or lack there-of) therefore your question is a deep one as Fellippe points out. I started down the AI path back in 1992 when I published a connect four program in a magazine that used an AI engine of my own design. I started reading books on the subject and soon realized the complexities involved in programming such beasts would need to be a full time endeavor with training in many disciplined areas.

My suggestion is to find books on the AI subjects you wish to explore, or better yet, find a forum where these deep subjects are discussed by other interested parties.

Good luck in your studies and I do look forward to any code you may share in the future. The subject fascinates me.

[bplus]

I too am a fan of AI but apparently I have a limited tolerance for abstract theory and prefer to perform experiments in code.

I too wrote up an AI for Connect 4, but not sure I want to publish. It's just coding how I would play only it can see diagonal connects way better than I, :) I have to be on my toes just to tie. I have read that there is a way to play a perfect game and if you go first you are a sure winner.

And then there was Battleship! One of these days I might get to chess, but I think checkers would have to be first.

So Dimster I hope you stay with this forum and keep us updated on your progress. The field is very broad from game AI to facial or voice recognition to driving and avatar's on Mars... do you have a particular area of interest?

BTW are you asking about training a Neural Net with weights and bias talk?

[Dimster]

Thanks Guys - the code I have written so far is rough, infantile and leads me to believe the computer (not me) is very dumb. Yes bplus I'm trying to understand the Sigmoid Neurons in terms of the basic language. Appears a lot of the present AI coding is in Python. Back on the .net I had some good advice on further reading for Decision Algorythms which has lead me to this point. I think I have a few more years of reading and coding before my SN can can raise its game to kindergarten status. But the joy is in the hunt. I think what I will try is both approaches for a BIAS.

[Bert22306]

Dimster, I think the answer is yes.

This is a philosophical question, and I'd say the answer depends on who wants the AI app. In my line of work, the customer expects predictable and repeatable performance, and not something you might use as a demo, to show of how AI makes decisions like haphazard human beings. So any bias has to be configured by the programmer, responding to performance requirements.

On the other hand, I might imagine a machine learning program, which determines outcomes for different "biases," and then adjusts accordingly over time. But as I alluded to before, the unpredictability of such a program would make it unacceptable in certain control applications. My 2 cents' worth.

[Dimster]

Thank you Bert

[SMcNeill]

I think all AIs have to start first with the programmer's bias, and then evolve from there.   Consider something like a predictive spellchecker: 

First, you'd program it with a dictionary.  Then you'd tell it what YOU thought the common words should be, and a few basic rules of grammer.  "Joe ATE dinner", instead of "Joe EAT dinner."  In this case, you're teaching it some basic tense structure....

And the user goes back and corrects the issue by making it, "Joe, EAT dinner".  Now instead of saying what Joe did in the past, it's a direct command for him to do something in the present...

The AI started with what you taught it, but it records the user's actual style of typing/writing, and in time it changes the response from "change EAT to ATE", to become, "ADD comma after subject".

It starts with what we -- the programmers -- know/think to be correct, and then it "learns" from there.

[bplus]

I watched this, basic Perceptron,

and think this might be translated to QB64.

Hi Dimster, Is this the kind of stuff you are looking into from Python video's?

When you talk of events and their probabilities it sounds like probability theory.
Isn't bias another name for weight?

Anyway in video link above, I learn that weights are started with Random values and are "trained" with computer program to match a given set of samples, improving the guessing process slowly, adjusting weights incrementally according to errors in guesses.

It all looks so easy until you go and try to duplicate what you saw. ;D

Append:
Oh here is the book author referred to in video link above:
https://natureofcode.com/book/chapter-10-neural-networks/
EDIT: Daniel Shiffman's book

Append #2: Oh there's the bias! used to prevent washout from multiplications by 0.

[TempodiBasic]

Hi Dimster
in my little experience I Agree with Steve

It starts with what we -- the programmers -- know/think to be correct, and then it "learns" from there.

in the way that if you create an AI and you give to AI: a memory ( a system to record) , an insight (a system to evaluate that is stored), a judgment ( a system to compare actual parameters in use with that comes up from stored experience), an autodrive (a system to change behaviores: divinding it in a system to create behavioures and a system to evaluate behavioures ) then you can give to AI a bias as addictional  behaviour or you can believe that in a long time of evolution (experience->evaluation ->adjourning parameters -> doing many alternative behavioures using new parameters ->evaluating behavioures-> select best behavioures to use)  the AI that you have coded evaluate the bias as a good value/parameter to bring with it.

So you must have strong flexibile algorythm to gain these results with an AI...
See  how rigid is the algorithm to do the best choice among all those possible.
https://www.hackerearth.com/blog/artificial-intelligence/minimax-algorithm-alpha-beta-pruning/

On the other hand I remember that on the offline website (.net) someone posted a link to a video in which we can see how AI can learn playing a videogame and how AI can increase its performances using neural network system.

Good Coding

[bplus]

OK it looks like Bias is treated as another input to "weigh" and adjust in training sessions, it has automatic value of 1 so when all the other inputs might be 0 there is still a value to weigh ie the sum won't be 0. Bias is a way to track the origin, could be very important for pattern recognition the relation to the origin,

[Pete]

I did an AI for a card game once. It would track the players bets and plays and form a bias on how it would respond based on that data. It seemed like it played better when a player played better, but it wasn't completely AI. Good play was based on a known set of playing statistics, so all the AI was really doing was adjusting its play to a randomized set of bets and moves within a skill level database, determined by the player's deviation from the best known possible bet and play.

Pete

- My wife says I have N.I. which I assume means "Natural" Intelligence, but she says that contains too many syllables.

[Dimster]

Thanks again all - I have read the articles/watched YouTube and all of it is very helpful. I did view that video back on .net but didn't keep the link.

[bplus]

OK Bias in action!

Code: QB64: [Select]

1. OPTION _EXPLICIT
2. _TITLE "NN 1" 'by B+ started 2018-09-08 my first attempt with training a Neural Net
3.  
4. '2018-09-09 retest this updating the Bias Weight with the others, though I suspect it will stir up a storm? Nope!
5.  
6. CONST WW = 800
7. CONST WH = 620
8. SCREEN _NEWIMAGE(WW, WH, 32)
9. _SCREENMOVE 300, 60
10.  
11. 'need a place to store inputs x, y, B for bias
12. 'well just use loops
13. 'dim and init weights
14. CONST LR = .2 'Learning Rate don't overcorrect errors
15. DIM SHARED BiasWeight AS SINGLE, Bias AS INTEGER, ArrayStart AS INTEGER, ArrayEnd AS INTEGER, SQ AS INTEGER, Mode AS INTEGER
16. DIM i AS INTEGER, rx AS INTEGER, ry AS INTEGER, colr AS LONG, cnt AS INTEGER, x AS INTEGER, y AS INTEGER, correct AS SINGLE
17. Mode = 1
18. ArrayStart = 1 'Start range of x, y
19. ArrayEnd = 60 'End range of x, y
20. SQ = 10 'for graphic squares drawing to watch Perceptron Learn Y > X
21.  
22. 'setup weights
23. DIM SHARED WX(ArrayStart TO ArrayEnd) AS SINGLE, WY(ArrayStart TO ArrayEnd) AS SINGLE
24. Bias = 1
25. BiasWeight = RND * 2 - 1
26. FOR i = ArrayStart TO ArrayEnd
27.     WX(i) = RND * 2 - 1
28.     WY(i) = RND * 2 - 1
29. NEXT
30.  
31. '61 x 61 squares of side 10 pixels fit in 610 x 610 pixel area
32. 'train 1000 random points per frame 10 frames per second and show progress of learning hopefully
33. WHILE _KEYDOWN(27) = 0
34.     CLS
35.     IF Mode = 1 THEN _TITLE "Training Pattern #1: Y >= X" ELSE _TITLE "Training Pattern #2: Framed!"
36.     cnt = 0
37.     FOR i = 1 TO 1000 'train 10% of points at a time
38.         rx = INT(RND * ArrayEnd) + 1: ry = INT(RND * ArrayEnd) + 1 '60 x 60 field so can show circles of radius 10
39.         Train rx, ry
40.     NEXT
41.     FOR y = ArrayStart TO ArrayEnd
42.         FOR x = ArrayStart TO ArrayEnd
43.             IF Mode = 1 THEN
44.                 IF TrainThis%(x, y) = -1 THEN LINE (x * SQ, y * SQ)-STEP(SQ, SQ), _RGB32(0, 0, 255), B
45.                 IF TrainThis%(x, y) = 1 THEN LINE (x * SQ, y * SQ)-STEP(SQ, SQ), _RGB32(255, 255, 255), B
46.                 IF TrainThis%(x, y) - Perceptron%(x, y) THEN 'wrong! Perceptron
47.                     colr = _RGB32(255, 0, 0)
48.                 ELSE 'good job Perceptron!
49.                     colr = _RGB32(0, 200, 0): cnt = cnt + 1
50.                 END IF
51.             ELSE
52.                 IF TrainThat%(x, y) = -1 THEN LINE (x * SQ, y * SQ)-STEP(SQ, SQ), _RGB32(0, 0, 255), B
53.                 IF TrainThat%(x, y) = 1 THEN LINE (x * SQ, y * SQ)-STEP(SQ, SQ), _RGB32(255, 255, 255), B
54.                 IF TrainThat%(x, y) - Perceptron%(x, y) THEN 'wrong! Perceptron
55.                     colr = _RGB32(255, 0, 0)
56.                 ELSE 'good job Perceptron!
57.                     colr = _RGB32(0, 200, 0): cnt = cnt + 1
58.                 END IF
59.             END IF
60.             LINE (x * SQ + 2, y * SQ + 2)-STEP(SQ - 3, SQ - 3), colr, BF
61.         NEXT
62.     NEXT
63.     correct = INT(cnt * 10000 / (61 * 61)) / 100
64.     _PRINTSTRING (650, 10), "Correct:" + STR$(correct) + "%"
65.     IF Mode = 1 THEN
66.         _PRINTSTRING (640, 30), "White Frame Y >= X"
67.         _PRINTSTRING (640, 50), " Blue Frame Y <  X"
68.     ELSE
69.         _PRINTSTRING (620, 30), "White Frame Train +1"
70.         _PRINTSTRING (620, 50), " Blue Frame Train -1"
71.         _PRINTSTRING (640, 180), "Wait for it... ;)"
72.     END IF
73.     _PRINTSTRING (640, 70), "Green Fill Correct"
74.     _PRINTSTRING (640, 90), "Red Fill Incorrect"
75.     _PRINTSTRING (620, 130), "Training 1000 random"
76.     _PRINTSTRING (620, 146), "points in 10th second."
77.     IF correct > 95.51 AND Mode = 1 THEN
78.         Mode = 2
79.         mBox "Hmm... there seems to be a real battle at the border. OK 95.5% has been exceeded, let's see how fast this retrains to a new pattern...", "OK Now Test New Training Set!"
80.     ELSEIF Mode = 2 AND correct > 96.5 THEN
81.         mBox "Over a 96.5% chance you're pregnant!", "OMG!"
82.         SYSTEM
83.     END IF
84.     _DISPLAY
85.     _LIMIT 10
86. WEND
87.  
88. FUNCTION Perceptron% (x AS INTEGER, y AS INTEGER)
89.     DIM sum AS SINGLE
90.     sum = x * WX(x) + y * WY(y) + Bias * BiasWeight 'sum the inputs times weight
91.     Perceptron% = Sign%(sum) 'apply the activation function to the sum for output
92. END FUNCTION
93.  
94. FUNCTION Sign% (n AS SINGLE) 'very simple activation function
95.     IF n < 0 THEN Sign% = -1 ELSE Sign% = 1
96. END FUNCTION
97.  
98. 'this sub trains one randomly chosen Perceptron weight set
99. SUB Train (rx AS INTEGER, ry AS INTEGER)
100.     'adjust Perceptrons weights until get good results
101.     '1. provide Perceptron with Inputs for which the correct answer is known
102.     '2. have perceptron guess the answer
103.     '3. Compute the error
104.     '4. Adjust weights according to error
105.     '5. repeat until shaped up
106.  
107.     'so what are we going to train , oh there it is
108.     DIM Guess AS INTEGER, Correct AS INTEGER, Errror AS SINGLE, OK$
109.     Guess = Perceptron%(rx, ry)
110.     IF Mode = 1 THEN
111.         Correct = TrainThis%(rx, ry)
112.     ELSE
113.         Correct = TrainThat%(rx, ry)
114.     END IF
115.     Errror = Correct - Guess 'either 0 when Guess = Correct  -2 or 2 when not
116.     IF Errror THEN
117.         WX(rx) = WX(rx) + rx * Errror * LR
118.         WY(ry) = WY(ry) + ry * Errror * LR
119.         BiasWeight = BiasWeight + 1 * Errror * LR
120.     END IF
121. END SUB
122.  
123. FUNCTION TrainThis% (x AS INTEGER, y AS INTEGER)
124.     IF y >= x THEN TrainThis% = 1 ELSE TrainThis% = -1 'the x = y function is the line between true and false
125. END FUNCTION
126.  
127. FUNCTION TrainThat% (x AS INTEGER, y AS INTEGER) 'draw a frame
128.     IF x = ArrayStart OR x = ArrayStart + 1 OR x = ArrayEnd - 1 OR x = ArrayEnd THEN
129.         TrainThat% = 1
130.     ELSEIF y = ArrayStart OR y = ArrayStart + 1 OR y = ArrayEnd - 1 OR y = ArrayEnd THEN
131.         TrainThat% = 1
132.     ELSEIF x >= 29 AND x <= 32 AND y >= 12 AND y <= 49 THEN
133.         TrainThat% = 1
134.     ELSEIF y >= 29 AND y <= 32 AND x >= 12 AND x <= 49 THEN
135.         TrainThat% = 1
136.     ELSE
137.         TrainThat% = -1
138.     END IF
139. END FUNCTION
140.  
141. 'title$ limit is 57 chars, all lines are 58 chars max
142. ' version bak 2018-09-07_10P
143. SUB mBox (m$, title$)
144.  
145.     'first screen dimensions items to restore at exit
146.     DIM sw AS INTEGER, sh AS INTEGER, fg AS _UNSIGNED LONG, bg AS _UNSIGNED LONG
147.     DIM curScrn AS LONG, backScrn AS LONG, mbx AS LONG 'some handles
148.     DIM ti AS INTEGER, limit AS INTEGER 'ti = text index for t$(), limit is number of chars per line
149.     DIM i AS INTEGER, j AS INTEGER, ff AS _BIT, add AS _BYTE 'index, flag and
150.     DIM bxH AS INTEGER, bxW AS INTEGER 'first as cells then as pixels
151.     DIM mb AS INTEGER, mx AS INTEGER, my AS INTEGER, mi AS INTEGER, grabx AS INTEGER, graby AS INTEGER
152.     DIM tlx AS INTEGER, tly AS INTEGER 'top left corner of message box
153.     DIM lastx AS INTEGER, lasty AS INTEGER, r AS INTEGER
154.     DIM b$, c$, tail$, d$
155.     sw = _WIDTH
156.     sh = _HEIGHT
157.     fg = _DEFAULTCOLOR
158.     bg = _BACKGROUNDCOLOR
159.     'screen snapshot
160.     curScrn = _DEST
161.     backScrn = _NEWIMAGE(sw, sh, 32)
162.     _PUTIMAGE , curScrn, backScrn
163.  
164.     'setup t$() to store strings with ti as index, linit 58 chars per line max, b$ is for build
165.     REDIM t$(0): ti = 0: limit = 58: b$ = ""
166.     FOR i = 1 TO LEN(m$)
167.         c$ = MID$(m$, i, 1)
168.         'are there any new line signals, CR, LF or both? take CRLF or LFCR as one break but dbl LF or CR means blank line
169.         SELECT CASE c$
170.             CASE CHR$(13) 'load line
171.                 IF MID$(m$, i + 1, 1) = CHR$(10) THEN i = i + 1
172.                 t$(ti) = b$: b$ = "": ti = ti + 1: REDIM _PRESERVE t$(ti)
173.             CASE CHR$(10)
174.                 IF MID$(m$, i + 1, 1) = CHR$(13) THEN i = i + 1
175.                 t$(ti) = b$: b$ = "": ti = ti + 1: REDIM _PRESERVE t$(ti)
176.             CASE ELSE
177.                 IF c$ = CHR$(9) THEN c$ = SPACE$(4): add = 4 ELSE add = 1
178.                 IF LEN(b$) + add > limit THEN
179.                     tail$ = "": ff = 0
180.                     FOR j = LEN(b$) TO 1 STEP -1 'backup until find a space, save the tail end for next line
181.                         d$ = MID$(b$, j, 1)
182.                         IF d$ = " " THEN
183.                             t$(ti) = MID$(b$, 1, j - 1): b$ = tail$ + c$: ti = ti + 1: REDIM _PRESERVE t$(ti)
184.                             ff = 1 'found space flag
185.                             EXIT FOR
186.                         ELSE
187.                             tail$ = d$ + tail$ 'the tail grows!
188.                         END IF
189.                     NEXT
190.                     IF ff = 0 THEN 'no break? OK
191.                         t$(ti) = b$: b$ = c$: ti = ti + 1: REDIM _PRESERVE t$(ti)
192.                     END IF
193.                 ELSE
194.                     b$ = b$ + c$ 'just keep building the line
195.                 END IF
196.         END SELECT
197.     NEXT
198.     t$(ti) = b$
199.     bxH = ti + 3: bxW = limit + 2
200.  
201.     'draw message box
202.     mbx = _NEWIMAGE(60 * 8, (bxH + 1) * 16, 32)
203.     _DEST mbx
204.     COLOR _RGB32(60, 40, 25), _RGB32(225, 225, 255)
205.     LOCATE 1, 1: PRINT LEFT$(SPACE$((bxW - LEN(title$) - 3) / 2) + title$ + SPACE$(bxW), bxW)
206.     COLOR _RGB32(225, 225, 255), _RGB32(200, 0, 0)
207.     LOCATE 1, bxW - 2: PRINT " X "
208.     COLOR _RGB32(60, 40, 25), _RGB32(255, 160, 90)
209.     LOCATE 2, 1: PRINT SPACE$(bxW);
210.     FOR r = 0 TO ti
211.         LOCATE 1 + r + 2, 1: PRINT LEFT$(" " + t$(r) + SPACE$(bxW), bxW);
212.     NEXT
213.     LOCATE 1 + bxH, 1: PRINT SPACE$(limit + 2);
214.  
215.     'now for the action
216.     _DEST curScrn
217.  
218.     'convert to pixels the top left corner of box at moment
219.     bxW = bxW * 8: bxH = bxH * 16
220.     tlx = (sw - bxW) / 2: tly = (sh - bxH) / 2
221.     lastx = tlx: lasty = tly
222.     'now allow user to move it around or just read it
223.     WHILE _KEYDOWN(27) = 0 AND _KEYDOWN(13) = 0 AND _KEYDOWN(32) = 0
224.         CLS
225.         _PUTIMAGE , backScrn
226.         _PUTIMAGE (tlx, tly), mbx, curScrn
227.         _DISPLAY
228.         WHILE _MOUSEINPUT: WEND
229.         mx = _MOUSEX: my = _MOUSEY: mb = _MOUSEBUTTON(1)
230.         IF mb THEN
231.             IF mx >= tlx AND mx <= tlx + bxW AND my >= tly AND my <= tly + 16 THEN 'mouse down on title bar
232.                 IF mx >= tlx + bxW - 24 THEN EXIT WHILE
233.                 grabx = mx - tlx: graby = my - tly
234.                 DO WHILE mb 'wait for release
235.                     mi = _MOUSEINPUT: mb = _MOUSEBUTTON(1)
236.                     mx = _MOUSEX: my = _MOUSEY
237.                     IF mx - grabx >= 0 AND mx - grabx <= sw - bxW AND my - graby >= 0 AND my - graby <= sh - bxH THEN
238.                         'attempt to speed up with less updates
239.                         IF ((lastx - (mx - grabx)) ^ 2 + (lasty - (my - graby)) ^ 2) ^ .5 > 10 THEN
240.                             tlx = mx - grabx: tly = my - graby
241.                             CLS
242.                             _PUTIMAGE , backScrn
243.                             _PUTIMAGE (tlx, tly), mbx, curScrn
244.                             lastx = tlx: lasty = tly
245.                             _DISPLAY
246.                         END IF
247.                     END IF
248.                     _LIMIT 400
249.                 LOOP
250.             END IF
251.         END IF
252.         _LIMIT 400
253.     WEND
254.     'put things back
255.     COLOR _RGB32(255, 255, 255), _RGB32(0, 0, 0): CLS
256.     _PUTIMAGE , backScrn
257.     _DISPLAY
258.     COLOR fg, bg
259.     _FREEIMAGE backScrn
260.     _FREEIMAGE mbx
261.     _AUTODISPLAY
262.     _KEYCLEAR
263. END SUB
264.  
265.  
266.  
267.  

Append: Oh crud! Was I suppose to update the BiasWeight also?
EDIT: this code edited since posting, I added update to BaisWeight but didn't make a big difference; it might have sped up the time to hit target correct %'s.

[Dimster]

So the difference between a "weight" and a "bias" in AI programming are very subtle bplus??? ie "Bias =1 : BiasWeight= RND*2-1". I do see the logic to your previous comment of bias simply being treated as another input to weight. If you were to decided an AI program MUST have an identifiable "Weight" and identifiable "Bias", would you say your "Bias = 1" would change to "Weight =1", as this value is more concrete and fixed. Or would you have "BiasWeight = RND*2-1", would in fact be considered as the "Weight". because weight in an AI program is always changing by the bias. Confusing I know. I guess, because of what I'm reading in AI programming, there is always a distinction between weight and bias.. Very much appreciate the comments in your program and the 96.5% chance of pregnancy.