Subscribe via Email

Saturday, May 21, 2016

morris_AIMA [#7]: Chapter 1 Solution Set [unofficial]

Here is my own solution set for Russel and Norvig's Artificial Intelligence: A Modern Approach. 

Disclaimer: use these solutions purely for personal enjoyment/comparison to your own. I completed this during my first semester of 2nd year computer science, so I take no claim to be insightful or intelligent on these matters. These are just my attempts to follow along with the book to the best of my ability. Hopefully I'll look back here in a few years and see how far I've come. 


(oct 7 2015) Intelligence: That in which allows an object/thing to make the rational, and optimal decision based on a pursued outcome.  
(oct 7 2015) Artificial Intelligence: The study of how computational agents replicate human, and general intelligence in order to think and act rationally. 
(oct 7 2015) Agent: That in which may perform an action. 
(oct 7 2015) Rationality: The quality of acting to pursue optimal outcomes based on logical interpretation of contained information.
(oct 7 2015)  Logical Reasoning: The act of making conclusions based on the formal properties and structures of logic. 


(oct 7 2015)

     Which objections still carry weight?
         "If the man were to try and pretend to be the machine he would clearly make a very poor showing. He would be given away at once by slowness and inaccuracy in arithmetic"
              The refutation to this argument claims that upon a machine being truly "intelligent" within this game, a man's flaws would not be enough to convince the interrogator of an optimal choice, as the machine would have a more "intelligent" way of convincing them otherwise. 
                   A possible issue with this refutation is that it begs the question, as in assuming that if the machine is intelligent, it will just find a way. That requires further definition of intelligence outside of simply defining this games properties. Turing avoids this by claiming he won't acknowledge intelligence outside of the scope of a typical man.
     Are his refutations valid?
(oct 9 2015)
          The Theological Objection:
              Thinking is a function of man' s immortal soul. God has given an immortal soul to every man an d woman, but not to any other animal or to machines. Hence, no machine can think"
              Turing refutes this by stating that we are not recreating the function of mans soul; but rather transferring it to another vessel. Also, many previous objections from theological perspective have been proven ridiculous after scientific advancements are made, so we should not be eager to use it to refute machine intelligence just yet. 
          The Heads in the Sand Objection
                   "consequences of machines thinking would be too dreadful. Let us hope an d believe that they cannot do so."
                    Turing states this argument is not sufficiently requiring refute, as much as those who deliver such an argument require comfort in their superiority. 
          The Mathematical Objection
              There are theorems that state any sufficiently complex logical system has itself statements in which can neither be proven nor disproved without showing the logical system itself is inconsistent. 
              In short, Turing argues that AI is the replicate human intelligence, and there is no confirmed proof that humans follow any different logical system from the one outlined in the objection. There may be things humans can/can't prove, and for that reason, it is arguable that the imitation takes into account for this. 
(oct 10 2015)
          The Argument from Consciousness 
              "Machines don't think unless they can feel the emotions humans do, like anger, love, remorse, etc"
                Turing refutes this by arguing that we wouldn't know if a machine feels this or not, just like we don't know this about any other human, rather we assume such in order to avoid a solipsist point of view. To refute one who uses this argument, Turing believes it would be easier to convince them to abandon their reasoning and agree with the imitation test, rather than be forced to agree with the solipsist point of view. 
          Argument from Various Disabilities 
          Arguments from Various Disabilities. These arguments take the form, " I grant you that you can make machines do all the things you have mentioned but you will never be able to make one do X "
          Turing argues this amounts to another form of the Argument from Consciousness, and that individuals claim disabilities of machines just as they refuse to believe anything can improve from a disability. Likewise, a machine learning like a human, with induction, may naturally make mistakes just as humans do, it is not the objective to remove all possible "mistakes".
         The Argument from Informality of Behaviour
              If man had a definite set of rules of conduct by which be regulated his life he would be no better than a machine. But there are no such rules, so men cannot be machines.' 
              Turing argues there may very well be such rules, despite them not being yet identified. 
        The Argument from Extra-Sensory Perception.
              "how will the imitation test work if stuff like telepathy exists"
              Turing argues that if something as ridiculous as telepathy exists, then a machine could be affected by such things as well, and the test would need to be ESP proof, by making a couple simple modifications. 
     Can you think of new objections arising from developments since he wrote the paper?
     In the paper, he predicts that, by the year 2000, a computer will have a 30% chance of passing a five-minute Turing Test with an unskilled interrogator. What chance do you think a computer would have today? In another 50 years? 
          At least 35% chance. Depends on whether or not people have gotten dumber or computers smarter.


(October 12 2015) I'd argue they are rational, in those obtained through evolution have demonstrated their ability to sustain life. Rationality is the direct result of evolution, where that in which survives, has made decisions based primarily on fact and reason from its past. For example reflexes to burning, is the response to the fact that fire is hot, heat burns, and being burned is bad, therefore it is in ones reasonable self-interest to not burn themselves. Following the dimensions of intelligence defined in this chapter, I believe that intelligence is defined largely by the the ability to act rationally, thus such reflexes are indeed intelligent; but how much so is a different question. 

1.4 (October 16 2015)

I believe the IQ test is originally designed to model the general intelligence of a human who, in fact has general intelligence. Designing an AI to perform exceptionally well on a specific IQ test doesn't necessarily mean it is more intelligent than a human, it just means the AI has better performance in a specific microworld, in which case many computers already do. For example, the average descent machine is eons ahead of humans in terms of arithmetic speed of calculation, but we don't say machines are more intelligent than humans because of that. Likewise, a machine can remember a lot more, and recall more accurately than a human (subject to argument); and the same scenario holds as mentioned before. 


(December 2 2015)
If we consider the number of memory updates per second relating to cycle time, and say
memory updates per second = #neurons * cycles per second, then with a sea slug consisting of 20,000 neurons, the memory updates per second would be 20,000 * (1000) = 2 * 10^7 then the current supercomputer has about 10^7 times more updates per second than the sea slug would. 


 (December 6 2015)

We all view the world in a different way, which may give bias to how we interpret information. I don't think one can be "wrong" about what they are thinking; however the action in which they use to portray and describe what they are thinking may be erroneous. When someone thinks, they are thinking, there is nothing fallacious about that; but the translation from thought to language, to the physical act of utterance or communicating may yield inaccuracy in information. Also on a really, really specific picky level, in order to process what you are thinking about, you must recall that information, but upon recalling that information you are thinking about something else, and not what you were recalling in the first place. 


(December 14 2015)
    Bar Code Scanners: Not really, strict algorithm with fixed in/output based on database of values.
    Web search engines: Yes, very much an AI. Given the environment (web) it may enhance performance based on your search queries, and give new suggestions based on them. As an agent, it responds changes in input (user queries) by adjusting it's search results, which may improve over time.
    Voice activated telephone menus: May have traces of this if it learns from input, depending on how well it may interpret a voice
    Internet routing algorithms: These do act rationally, in that given the input (state) they may act in the most efficient manner. 


(December 14 2015):
    The reason we cannot "do" it may be more due to how we physically communicate via mathematical models and forms of language. Our brain may have a more optimized solution that can still perceive the same structure; but is using a different operation. For example in abstract algebra multiple transformations may be expressible using different basis of vector space; yet one of those such basis has a much more simple (diagonalized) matrix for the transformation, but both matrices represent the same inherent transformation with respect to the vector space.


(December 16 2015): One of the main goals for humanity is survival, and since we consider evolution to follow the survival of the fittest principle, the act of surviving is rational, so evolution leads to more rational agents. 


Based on definition of engineering "the branch of science and technology concerned with the design, building, and use of engines, machines, and structures.", AI falls into this category.


(December 26, 2015)
If programmers are intelligent, though, why shouldn't they be able to program intelligence? Intelligence is something growing, once the foundation is set further intelligence can be gained from an agent. We would need to understand ourselves better before I believe we can implement it though. 


(December 29, 2015): I lack a foundational in biology so this is difficult to answer. I'm tempted to argue that genes aren't the only thing that regulates an animals behaviour and that there are other, perhaps internal factors that effect how an animal may interact intelligently with its environment. 


I do agree with the latter statement, in that physical laws govern all actions. The issue with this implying the former is that on a quantum level for example; much randomness occurs despite these laws, and even if we are controlled by something else, we have defined "intelligence" with respect to our own existence and experience. By this I mean when humans defined intelligence they looked at how an intelligent human interacts and WE defined it in that way regardless of physics, thus I believe both humans and machine/animals as well may become intelligent under the same argument that we defined intelligence without consideration of the deep-rooted physics on it; but rather based on the ability to act rationally, however we as a race deem that to be. 


    a) Dr. Peters used human demonstration/learning techniques and reinforcement learning to teach a robot to hit ping pong balls into a cup, and bounce in place. (2010). The KUKA robot group published an over dramatised video showing a demo of ping pong against a human; but behind the scenes details revealed there were many flaws. It does not appear this game is "solved" by any means.
    b) Not currently, too unpredictable for current algorithms/vision techniques
    c) Yes, during DARPA's final challenge, 6/11 teams were able to finish the course manoeuvring through the streets in various environments. Focus is needed on driver/vehicle interaction since the driver has no role any more, also focus on diagnostics, as in the vehicle needs to know when it is not safe/able to drive, and from a human standpoint validation is needed to ensure what constitutes a safe autonomous vehicle. source

    d) Tally is a robot that helps with autonomous product management in grocery stores by letting managers know when products are low on quantity/priced wrong. There are other agents that may assist in helping you shop, and deliver groceries; but none in particular that actually selects the products for you, as that is likely just personal preference more than intelligence. 

    e) There are many recommendation systems that help in providing feedback on what you may like, so this is in the machine learning/data mining domain. Again similar to part d, there does not seem to be a focus on developing something that actually makes the choices for you. 

    f) Theoretically, the game is solved is there was a fast enough computer to process a massive game tree of all possible moves; but that's not possible at this time so the problem is technically not solved. There is a technique using hierarchical task tree planning to eliminate the overhead a bit (source) however it still crumbles in comparison to even decent players at a local lounge. 

    g) This is a whole branch of AI. There is lots of research on it, high-level symbolic logic tools and theorem proving assistants; but no definite fully complete theorem prover. Of course, if such a thing existed, we'd know due to the impact it'd have on the mathematics community. 

    h) There exist tools to generate simple puns/jokes such as the "JAPE" (joke analysis and production engine).

    i) CCLIPS was worked on around 1986 with focus on being a rule-based system that interprets legal information and determines the effects of certain decisions. Ross, a recent addition/rendition of Watson is a cognitive computer that provides relevant answers to legal questions and aids in helping lawyers make efficient decisions. At this time there doesn't seem to be a machine that has "solved" the need for a lawyer, as the final decision is often left on the moral side despite rigorous legalities. 

    j) Big improvements on this as of lately; but still not "solved". The BCI was a prototype machine-translation system from 1992 that did a transfer of language using Quasi Logical Form; but of course was not full proof. Google translate and many other apps provide quick translation so this can be considered "solved" but there is always room for improvement. 

    k) Not even close; but some machines can assist with steady movements, acting just as machine. We have not become confident enough in a machine to take on a task fully autonomously yet. There do however exist intelligent systems for guidance/performing small subtasks; but nothing to the scale this question is suggesting. 
Main difficulties in general involve how to interact with humans and shift over society to such a change, as long with how to model the perceptual input efficiently, and minimize uncertainty in data. For example with morals/emotion, machines have difficulty and rely on probabilistic techniques rather than fact/algorithms.


(December 30, 2015)
Contest choice 1: DARPA Grand Challenge for robotic cars
    As of 2005, it was a smashing success, with several vehicles completing the course, and signaling the age of autonomous vehicles. The competitions have a militaristic agenda which may draw away from other societal issues say around autonomous vehicles for the disabled. In the long run, it would appear everyone will get their own benefit from this competition; however as said the military will be the first, and as a result may indirectly damage innocent lives through military endeavours. 
 Contest choice 2: International planning competition:Starting in 1998 with 4 types of planning "Classical planning", "Hierarchical planning", "Reactive planning", and "Learning in planning". The competition was created due to the seemingly minimal usage of current search-based planning algorithms and a need to improve them. This contest helps advance the efficiency of many planning algorithms thus benefiting other fields of AI relying on them. Of course like any other contest, it draws attention away from perhaps more effective approaches to a problem. For example, it may be more desirable to focus on a deep-learning approach to abstraction ( as mentioned by demis hassibis 2015), rather than just a straight algorithm. 
    Contest choice 3: Robocup soccer: Big fan of this one due to aalab I work in. It promotes science/research among beginners and advanced researchers alike. The contest uses an evolution/revolution epoch approach, meaning there are set changes that take place of a specified epoch that promotes evolution of the approaches used and new techniques that advance the technology, where as often a "revolution" takes place as well launching into a new epoch with quite different rules to accompany the enhanced method. The competition may prevent research on other specific events since it is soccer focused; but really I don't see it as an issue considering how many domains are being covered with autonomous soccer. 
    Contest choice 4: TREC information retrieval event: focuses on various algorithms/approaches to collecting/presenting information. It increases communication between academics and the real world showing what has been achieved in the field and helps place the research into commercial products by showing the real-world application of such research. The problems are very specific so may neglect other topics; but any approach can be used to tackle them so the choices are wide.
    Contest choice 5: Kaggle: Kaggle and other similar sites provide data challenges and allow users to submit various scripts that may tackle them. It is beneficial as it helps with the practising implementation of algorithms and outsources company projects to the open source community. These problems are often very restricted in nature and lacking the real-world error and uncertainty and with a solution already existing to mark submission against, the contests aren't necessarily solving new problems but perhaps introducing a new solution to an older problem, which can be beneficial. 

No comments:

Post a Comment

Please feel free to give feedback/criticism, or other suggestions! (be gentle)

Subscribe to Updates via Email