|Assignment 2 - Kartik Vishwanath
Steps towards Artificial Intelligence – Marvin Minsky
Broken Glass Summaries
Steps towards Artificial Intelligence- Marvin Minsky
Heuristic programming is a method of incorporating intelligence into machines. Heuristics refer to “rules of thumb” that increase the efficiency of a machine in solving a complex problem, a heuristic may therefore fail in certain cases but overall can be assumed to reduce the complexity of finding the solution to problem considerably.
This paper discusses the areas into which heuristic programming can be divided namely Search, Pattern recognition, Learning, Planning and Induction.
Search: This refers to the technique of searching a solution space for a suitable solution to a problem. This assumes that the problem is well defined in the sense that we can decide when a solution found out is acceptable. All these problems can be theoretically solved by simply exhausting all possibilities, but generally this method is inefficient for practical purposes. We can reduce the complexity of the search by techniques such as Hill climbing. In Hill climbing at attempt is made to maximize the output function by adjusting the input parameters. To do so we need to explore locally about a point and try to find out which direction in the space leads to the steepest uphill climb. This still suffers from many problems for instance if we reach a local maxima then we might incorrectly decide that it is the optimal solution. Another more serious problem is that of the “Mesa Phenomenon” in which locally moving in any direction leads to no performance improvement.
Pattern Recognition: This deals with the classification of problems into classes. Within a class a particular method of problem solving is most applicable. This acts as the defining feature for that class. For this the Heuristic features of the problem have to be identified and the method most suited for them has to be figured out.
Learning: The basic goal of this area is to apply similar methods to a particular problem that have worked on similar problems in the past. One method to achieve this is using success-reinforced models wherein success in the application of a particular method leads to a reward that increases the probability of that particular method to be used again in problems of a similar nature. For complex problems we can divide them into smaller problems each with local goals and methods taken to reach these goals can be reinforced.
Planning: While solving a problem we may be faced with a large number of sub problems. At each stage of problem solving a subset of problems has to be selected based on the relative difficulty and appropriate methods have to be selected to solve them. In cases of difficult problems the steps taken to search for selecting suitable sub problems and appropriate methods for them may become too difficult therefore we must look at the problem as a whole and analyze its structure, this is termed as planning.
Induction: This final area explores what intelligence actually defines. According to Minsky any machine that can generate a set structures (by learning) to solve a problem, that are not anticipated by the programmer can be said to be intelligent. Inductive inference deals with how machines can generate general explanations about events that are totally different from what it has ‘learned’ from experience. If a system can answer a hypothetical question without performing an experiment then it must have created an internal model of the world to represent that event. This internal model must represent the machine also thus creating self models that aid in introspection. Minsky believes that such self models will exhibit a similar kind of duality that human’s exhibit.
Computing machinery and Intelligence – A. M Turing
Turing attempts to answer the question “Can machines think?” He transforms the question using an “imitation game”.
This game is played among three people a man (A), a woman (B) and an interrogator (C). The Interrogator is apart from the other two. The object of the game is for the interrogator to identify which is the man and which is the woman in the end of the game. To find this out the interrogator is allowed to pose questions to both. The object of A is to mislead the interrogator (C) and try and cause him to make the wrong identification. The object of B is to help the interrogator (C) in making the right decision.
To explain his question transformation Turing assumes that a machine takes the part of A. He now poses the questions “Will the interrogator decide wrongly as often when the game is played like this as he does when the game is played between a man and a woman?” these questions replace the question “Can machines think?”.
Turing also does not attach much importance to visual, linguistic or motor capabilities. In his view they are just means of communication that can be replaced by other means. He believes that education can be imparted through any medium and does not necessarily require all the three capabilities mentioned above.
Similarities and differences in the views of Turing and Minsky
Both Turing and Minsky believe that machines can have intelligence and both support the view that the argument that machines can only do what they are instructed i.e., follow only a step by step procedure is not true. They argue this by explaining that in learning systems where a machine learns from its past experience in solving a problem the programmer of the machine can not predetermine the structures used by the machine in reaching the solution. Thus both of them share the view that in such a case machines can be assumed to posses intelligence.
The differences in their approach is that Turing approaches the question of whether machines can think and proposes a test to determine whether a machine can be deemed as an intelligent machine. Minsky on the other hand proposes the areas and the techniques to be used in those areas, using which a machine can be made to exhibit thinking or intelligent behavior. Turing’s conclusion is that any machine that can even imitate the thinking process and can successfully imitate a human in the his test is intelligent while Minsky relates the concept of intelligence to the internal model of a machine of itself. Minsky argues that a machine is truly intelligent only when it exhibits a duality in the interpretation of its own model much in the same way as humans do. Turing also has a less mathematical approach to the problem as compared to Minsky who tries to provide a mathematical backing to his theory.
Blocks world problem
The blocks world problem consists of forming a model of a concept. Checking this with respect to previously defined models of the same concept and merging it with them i.e. forming a generalization of the concept. Identifying models that are similar but belong to different concepts, and restricting the definitions of these concepts to exclude the models that nearly miss. Minsky’s discussion on learning and its techniques can be applied to this problem as well. The formation of a correct concept definition can be reinforced and thus over a period of time a more accurate model of the concept will get built up. Similar models of different concepts can be discovered by the process of reinforcement by the level to which a model is reinforced compared to actual models of that concept. Using this information we can define restrictions to concept definitions. Thus in a way Minsky theory of learning does anticipate the Blocks world problem.
Strengths and Weaknesses of Misky’s Conclusion
The strength of Minsky’s idea is that it clearly identifies areas into which research may potentially yield an artificially intelligent system in the future. Minsky also proves that problem solving itself encompasses almost all that we associate with intelligence and a good problem solving system can be quite intelligent.
The weakness of Minsky’s solution is that is focuses on the aspect of problem solving as the only aspect of intelligence other factors such as linguistic capabilities, motor capabilities that can also be linked to the concept of intelligence are largely ignored. Also Minsky does not elaborate on his philosophical concept of a self model of a machine. For instance how can a machine evaluate its own self model? A human being can view an event, question its occurrence and proceed to investigate why it occurred by forming a hypothesis. Minsky does mention a mention a machines capability to hypothesize but can a machine question the occurrence of an event on its own?
My own conclusion would be that to create artificial intelligence we require research primarily into 5 areas Search, Pattern recognition, Learning, Planning and Induction. We should incorporate other factors such as visual representations, linguistics etc. into the learning phase of an intelligent system.