AI is not pure science. There must be a mind in AI. Can the AI mind become sentient? There are a lot of philosophical conundrums to deal with. Is human consciousness computational?

The Turing test
If an individual cannot discern if he is dealing with a fellow human or a computer, then the computer passes the Turing Test, meaning the computer already thinks and engages at the level of a human.

One problem: does being human-like mean you are intelligent?
Eugene, a child chatbot answered unintelligible answers in the same way a short attentionspan kid would, but it fooled many judges and passed the Turing Test. So, maybe the Turing Test was not about intelligence but human behavior?

The Chinese room argument
Even though the guy outside received intelligent answers and thought he was dealing with a human being on the other side of the room, the other human being was only getting instructions from a book without really knowing what is being said. So, the 'Turing' computer may give all the right answers and fool the other guy, but there is no real intelligence at play - just broken down small mechanical instructions that can be automated.

Is a self-driving car intelligent?
When a car drives by itself, it appears intelligent because it is behaving intelligently. But intelligent behavior is not necessarily being intelligent the way a human is. AI systems do not understand its environment the way a human knows it's embroiled in traffic at rush hour. Being intelligent (human) is different from acting intelligently (self-sdriving car).

How much does philosophy matter in practice?
Philosophical discourse with friends over a few drinks is intellectually stimulating, but hardly makes a dent on the way AI systems are developed and manufactured. Philosophy and practice do not necessarily impact the other. Thus the focus of this course is the 'practical practice' part and not on philosophy.

2a. Search and Problem Solving

Search
Most problems are actually Search problems even if it's not too obvious at the outset. After identifying the problem, best way is to start with choices available and then the goal (at what point is the problem solved?). From those 2 end points, we work on the sequence that would take us from the problem to the end goal.

Toy problem: chicken crossing
It's a simple problem and it's easy to see the solution without setting up the 64 possible scenarios. But for complex problems with combinations running into millions, it's best to use the computational approach since human thinking will fail at that load.

State Space
This is the state of all possible situations to get from point A to the end goal. In the chicken-crossing puzzle, the state space consisted of ten allowed states NNNN through to FFFF (but not for example NFFF, which the puzzle rules don’t allow). If the task is to navigate from place A to place B, the state space could be the set of locations defined by their (x,y) coordinates that can be reached from the starting point A. Or we could use a constrained set of locations, for example, different street addresses so that the number of possible states is limited.

Transitions
This is the movement from one state to another. A series of movements constitute a Path.

Costs
This is the cost (not necessarily money) for the transition. Some transitions are cheaper than others. If the goal is to take the shortest distance, then kilometers would be the natural cost. If the goal is about getting there the fastest time, then time is the natural cost.

Exercise 6: The Towers of Hanoi

Below are all the possible states:

2b. Solving Problems with AI

Modern computer
Alan Turing maintained that anything computational can be automated. This led to the invention of the modern computer. During its time, it was ground-breaking and led in large measure to the defeat of Germany during WWII.

Intelligence is computational
John McCarthy, known as the Father of Modern AI, coined the term "artificial intelligence". He maintained that any aspect of learning can be quantized so that a machine can simulate it. In short, any element of intelligence can be broken down into simple mechanical steps that it can be written down as a computer program. Intelligence is intelligence, even if it came from a computer program. Intelligence is not necessarily a biological monopoly.

Why search and games became central in AI research
Subsequent development of computers gave rise to AI algoriths that were experimented on. The most popular field was in games - chess, checkers and ultimately, GO. Sophistication in Games, Search and Planning gave rise to advanced algoriths like Minimax Algorithm and Alpha-Beta Pruning.

2c. Search and Games

Tic-tac-toe Game trees
This is a simple 2-player game and a great example of a classic AI problem. Game tree is very similar to the Search state space diagram except that it starts from all the possible moves the opposing player can make. The game is over if someone makes a straight 3-in-a-row or if the game is a draw.

Minimizing and maximizing value
A numerical value is attached to each possible end result, e.g. "+1" when Max wins and "-1" when Min wins. It is "0" when it's a draw. When forced moves for an inevitable win is already evident (just like chess), then AI already knows the outcome long before the final move is taken. So, if Max will inevitably win, then AI gives a "+1" value even if the game hasn't officially played out in its entirety.

The Minimax algorithm
Minimax is a recursive algorithm that guarantees the optimal moves in a game for a sure win, if the outcome has already been determined in advance based on current positions and forced moves or inevitable win. It's goal is to make the best possible move from all possible moves given the predictable outcome of every move.

The following conditions must be met:

1. 2-player game - the variables are limited but the game tree can be huge like chess - for every 2 moves, the 'children' (possible moves) are about 1225! Three moves? Over 42k nodes (possible resulting positions). But even for just 2 people playing GO, the game tree is so vast that Minimax is no longer viable
2. deterministic - predictable outcome like moving a chess piece, no randomness like throwing a dice
3. zero-sum game - the other's gain is somebody's loss. There is no such thing as everybody wins...someone has to lose
4. perfect information - both players have complete knowledge of the game situation like chess...unlike when you throw a dice or draw from a shuffled card

More tricks: Managing massive game trees
For massive Game Trees, there are a few tricks, some of them employed with Deep Blue in defeating Kasparov in 1997. To stop the Minimax before it reaches the end game, the heuristic evaluation function is utilized given the current board game situation.

Good heuristics
Good heuristics assign a value to the weight of the chess pieces, e.g. a Queen has 9X the value of a pawn. It also weighs the remaining pieces on the board and where they are located.

Depth-limit
The number of steps that the Game Tree has been expanded to present a heuristic evaluation function at all nodes.

The limitations of plain search
The 'search' algorithm seems to be perfect and mechanical by simply going down the game tree until the optimized route is established. But that is not so simple in the real world. Even in a simple real-world situation, the game states can be so complex that heuristics and alpha-beta pruning cannot work anymore. Furthermore, the transition states may not be deterministic, out of our control or simply unknowable.

Chapter 3 introduces the added complexity of uncertainty and probability to simulate real world problems.

3a. Odds and probability

On paper, it's easy to have a car go from Point A to Point B while in compliance to all the laws. But in the real world, there is traffic, accidents, bad weather, pedestrians, etc. It's not so simple anymore. A self-driving car also has sensors (for non-visuals like sonar, infra red) and cameras (for visuals). They usually create noise which corrupts the data.

Probability
One of the reasons why modern AI methods actually work in the real world - as opposed to most of the earlier good old-fashioned methods in the 1960-1980s - is the ability to deal with uncertainty. The real world is full of uncertainty, missing information, and even deception.

Back in the day, there were many approaches to handling uncertainty in developing AI systems. The most common then was Fuzzy Logic. But now, it is unanimously accepted and practiced to use Probabilities to handle uncertainties.

Why probability matters
We are familiar with probabilities. What is the probability that tossing a coin will land heads up (50%). But there are other real-life application of probabilities too. What's the probability of rain in this date on the calendar? What's the probability that interest rates will go up in the next quarter? These are all uncertainties that we find an educated guess for answers.

The best way to understand probability is to see it as a quantifiable number, and therefore, comparable to other numbers. Probability provides a systematic way to quantify uncertainties so that it falls within the realm of rational thinking...and not gut-feel speculation.

Why quantifying uncertainty matters
If we think that uncertainty cannot be quantified or measured, then there is no basis for rational decision-making. And in the absence of that, making a decision becomes a hit-or-miss draw of the lot.

Since this is an AI course, automating uncertainty through probabilities is underscored, e.g. fitering spam.

Odds
Uncertainty is best understood by presenting the odds - this can more readily shift public assumptions. Odds are usually expressed as ratios, e.g. 3:1 (the ratio of milk to water is 3 parts milk and 1 part water)

Why we use odds and not percentages
75%, .75 and 3:1 all represent the same thing. But people get confused with percentages and fractions, so it's best to use odds and ratios in AI talk.

Odds (ratio) and Probabilities (% or fraction) are NOT the same
1:5 ratio is not the same as 20%. 1:5 loss means for every 6 games, there is one loss. 20% loss means for every 5 games, there is one loss. Keep this in mind when dealing with probabilities and odds.

Converting Odds to Probabilities
x:y ratio = x/(x + y)
5:1 ratio = 5/(5 + 1) = 5/6 (fraction) = 83% probability

3b. The Bayes rule

Updating with new evidence
The details and complexity of Probability Calculus will not be discussed in this course, but one formula must be taken up - the Bayes Rule or Bayes Formula or Bayesian Inference.

This formula is both simple and elegant. It is also very powerful and can be applied to nearly all of science in weighing conflicting evidence, in a court of law, spam filtering, medical diagnosis, etc. Of course, in AI, Bayes Rule is applied to Machine Learning, Data science and decision-making processes. Bayes rule is to update current probabilities based on new information.

Bayes Rule is expressed in many ways, one of which is Odds. We first take the odds of something happening (vs the odds of it NOT happening), the question being, "Given the new evidence, it the original hypothesis (Prior Odds) still true?".

The equation to compute for the Bayes Rule is:

H = (H), Hypothesis, the odds before new Evidence was obtained
E = (E), Evidence, new information that tests if the Prior (H) still holds true, E = E|H X H + E|*H X *H (*H is the negation of the H)
L = (E|H), Likelihood, odds that the old Hypothesis (H) still holds true given the new Evidence (E)
Posterior = Likelihood Ratio / Prior (H) = (H|E)

Prior odds
The Prior Odds - the odds is the current probability before new information became available. The idea is to update the Prior Odds when new information becomes available.

Posterior odds
New probability when additional information is added to the Prior odds. This is mathematically arrived at by: Posterior Odds = Prior Odds X Likelihood Ratio.

How odds change
New evidence or new information changes the odds. This new information adds a 'Likelihood ratio' which changes the Prior Odds to a new Posterior Odds.

Likelihood ratio
This is a measure given to determine how new information changes the odds of something remaining true. This ratio is essential in updating the Prior Odds to obtain the Posterior Odds. To get this result, the Bayes Rule has to be applied.

Base-Rate Fallacy
This is when people disregard the Prior Odds (base rate) and only look at the new information in making decisions. This is like making a decision based on a snapshot instead of seeing a trajectory or pattern from past events. This mistake happens because factoring-in the base rate requires more cognitive effort and statistical understanding which is daunting. Thus wrong conclusions are made. By using the Bayes Rule, this mis-step is avoided.

3c. Naive Bayes classification

One of the most useful applications of the Bayes rule is the so-called Naive Bayes classifier.

xxxxxxxxxxxxx

--- Gigit (TheLoneRider)
YOGA by Gigit | Learn English | Travel like a Nomad | Donation Bank

Next story:

Moving to Taloto, Tagbilaran City

(May 19, 2024) Even though I could have stayed at Kat's place until June 21, it was best to start looking for a suitable residence early on. So when I found this place with a bedroom, bathroom and kitchen/dining area just 2km away from Kat, I took it. At P6k, this doubles my budgeted rent and cuts in half my one year safety lifeline. But having this place already eased my racing mind and bought me time....more »»

»» back to Technology
»» back to Learning
»» back to Homepage

ARCHIVE: 2024 | JAN | FEB | MAR | APR | MAY | JUN | JUL

1970 | 1973 | 1975 | 1976 | 1979 | 1981 | 1996 | 2000 | 2001 | 2002 | 2003 | 2004 | 2005 | 2006 | 2007 | 2008 | 2009 | 2010 | 2011 | 2012 | 2013 | 2014 | 2015 | 2016 | 2017 | 2018 | 2019 | 2020 | 2021 | 2022 | 2023 | 2024 | ALL BLOGS