PROBLEMS IN AI IN 1975 AND RELEVANCE OF THESE PROBLEMS IN TODAY’S SCENARIO

ABSTRACT

A significant resurgence is seen in AI technologies over the years. AI has become commonplace in every aspect of life. The AI vision was put into words over seventy year ago when Alan Turing proposed the Turing test as measure of machine intelligence and Isaac Asimov published his Three Laws of Robotics. Artificial intelligence (AI) since then has been defined as method to put human knowledge into machine. The notion of what is an AI task has changed over the years, as problems once thought to be easy have turned out to be hard, and vice versa.

FEATUES OF INTELLIGENCE

There is list of features which we expect that intelligent entity must have. To analyze the intelligence five such features are communication, internal knowledge, world intelligence, intentionally, creativity.

· Communication: it is an important feature. We seek for an advice from an intelligent entity. It is one of the significant features of intelligence.

· Internal Knowledge: intelligent entity should not give partial or irrational solutions mean solutions do not possess any reasonable or practical significance. For this entity must have an internal knowledge.

· Creativity: there must be creativity in intelligent entity which means that they must be creative in finding the new paths and also plan from past experience in order to achieve future goals.

· Intentionality: Goal-driven behavior is another key feature of intelligence. Intelligent entities know what they want and how to get it. The key to intelligent intentionality is how appropriately and flexibly plans can be applied to novel situations.

· World knowledge: entity must have the experience and knowledge of outside world. Intelligent entities learn about the world from analyzing, observing and experiencing.

People consider AI problem of two types — solved and unsolved.

PROBLEMS IN AI IN 1975 AND RELEVANCE OF THESE PROBLEMS IN TODAY’S SCENARIO

PROBLEM: Representation

There is a need for some kind of machine for representing the knowledge to solve a real-life problem.

In earlier times, computer should work on ideas, faces and concepts. This was the point of representation. This scheme should be away from problems like metaphor, ambiguity, ellipsis. No important data similarity should be hidden and no such extra similarity should be added in case of textual representation.

Certain ways were discussed on how one could manipulate the knowledge with the help of a computer. One way was to represent data in the form of English text that could describe the fact or an idea. But it turned out to be tough nut for the computing machine.

The other mode of representation was conceptual dependency. CD wasan attempt to represent schematics of action in a clear and unambiguous form. But, CD was not able to handle complex action. So, more complex structures were needed.

Further, a couple of modes of representing knowledge like Memory Organizational Packets (MOPs), Explanation Patterns (XP) were discussed but they didn’t prove to be reliable as they had their drawbacks.

Presently, Different kinds of knowledge that need to be represented in AI include objects, events, facts, meta-knowledge and knowledge base. Different types of knowledge include declarative, structural, procedural, Meta and heuristic. Now, techniques of knowledge representation include logical representation, semantic net, production rules and frame representation. It is still considered as one of issues in AI and researchers are working on it. This could be one of the biggest achievements in the field of AI.

PROBLEM: Decoding

Representations of some statements are quite similar. For example,

The statements are as follows:

Person x took a bus.

Person x took a medicine.

Person x took Person y to a party.

The word ‘took’ has been used in all the three statements but the problem lies on what grounds the computer can decode the difference between these statements.

To distinguish, the machine could use some keyword. Like, in first statement word ‘bus’ represents that the sentence is related to travelling. The computer needs to know the things happening in real world. This makes the decoding easy.

Decoding was understood as a method to derive information from other sources like sensors. It was used for the interpretation of very similar things into dissimilar internal representation. There was a need of procedural knowledge associated with data in 1970s. Procedures were also known as Requests. Decoder needs to access full range of these requests to solve this problem completely. In 1970s solution in trend was MOPTRANS Parser. The principal problems in decoding are how memory can be arranged so that the correct knowledge is brought to bear at the correct time during the decoding process, and how the complexity of multiple, interacting, possibly conflicting requests or interpretations can be controlled.

The issues faced while decoding is the arrangement of memory when something is needed from it at a particular point of time during the decoding process, with the problems of multiple, interacting, possibly conflicting requests are tackled.

PROBLEM: Inference

This problem discusses the issues of drawing inference from the set of problem statements that are given to a computer. This problem arises just after the problem of decoding has been corrected.

An example has been discussed in this given problem and its interpretation has been talked about. The author takes the following example:

‘John took a bus to New York.’

The above sentence can have the following inferences:

John is a bus driver; john is a passenger; John lives in New York; john is poor; john wants to go to New York and so on…

Many different inferences can be drawn from the above statement. So, different results could increase the complexity of the given statement.

Inference is generating the conclusions from evidence and facts. This problem can be defined as how well our machine decides which inference to make and how to decide how far to extend the inference chain. Inference rules and pattern matching were the solution to this problem. It proposes that if the pre-condition is matched then machine should generate the inference.

Presently, the inference rules are being followed. Modulus Ponens and resolution are some of them. Modulus Ponens defines that if P is true and P → Q is true, then we can infer that Q will be true. The Resolution rule state that if P∨Q and ¬ P∧R is true, then Q∨R will also be true.

PROBLEM: Control of combinatorial explosion

Control of combinatorial explosion was considered as major challenge in the AI domain. This problem states that when we try to consider the inference chains of any length greater than a certain number of steps then the Combinatorial Complexity of trying to investigate such chains is humongous.

This problem talks about the number of inferences that should be included in the system’s memory. A number of inferences can be drawn from a given fact or previous inferences. It is difficult to process all because of memory constraints.

The author has focused on the solution that combining the general way to the solution of the problem with a solution of the problem of representations being too low level.

All the related task should be combined and should fall under the super event. An example has been given that a person in restaurant performs several action like ‘entering the restaurant’, ‘ordering’, ‘eating the food’, etc. According to author these events should fall under the super event ‘visit to restaurant. Basically, the concept of CD is used.

So, the combinatorial explosion results in limited size of problems that can be solved with brute force search techniques. Their time complexities grow exponentially with the problem size. Search techniques to contain combinatorial explosion in AI are uninformed or heuristic methods.

PROBLEM: Indexing

Indexingdescribes how data is organized in a storage system to facilitate information retrieval. The author focuses on how to index the problems efficiently and come up with events that are correctly indexed. Memory organization is considered very important in order to retrieve indexed patterns whenever required.

Making a model from scratch requires more processing the doing analytic work and retrieving it later. As one retrieves the complex models and other knowledge structures, it would be efficient to index these structures to retrieve the right structure at the right time. But if the size of the structure is very large making a model from the scratch could be a faster way.

A lot of processing and inferential power is required to perform a abstract task. It makes indexing a difficult task. Example given by author explains that John wants to make a call and he doesn’t have change so he goes to diner. But because of his location one can say that he is going to have meal which is not a case here.

This problem is still faced today to some extent as in order to cover all the possibilities and the vast user database that we work with these days’ memory indexing is required to increase the performance of the Ai system but it comes with its own drawbacks.

PROBLEM: Prediction and Recovery

The author discusses the problem of prediction and recovery. Sometimes the predictions made about the output are wrong concerning the input. Understanding of knowledge structure is important to make good set of predictions. It is expected from the machines to come up with the solution at the times of failure.

Predictions are not always correct. Sometimes they are wrong because the under stander has misunderstood what is actually happening and has activated the wrong knowledge. Or, the right knowledge structure is activated but some bizarre event has occurred.

Failures are classified as predictive failures, plan executive failures and money retrieval failure. All of them are different from each other.

One solution of recovery discussed by author is backtracking. But it was a non-theory as combinatories of this approach are impossible. It is not executable in real time as robot once walked off cannot backtrack and try its next option. Failure gives tremendous amount of information and system must learn from the failure.

As a failure occurs it is advised to use a different knowledge structure which could help in easy recovery from that failure. If one plan fails, the system should g for other plan to make recovery and solve the problem.

PROBLEM: Dynamic modification

This problem deals with the dynamic modification. We know that we go deep inside the topic before coming to a solution. Similarly, machine should also learn from knowledge structure and experience before coming to a solution.

The properties of a good model are that it will be able to make some good amount of explanations by referring reminding, or to froze patterns of casual reasoning. The author has referred to these casual patterns as Explanation Patterns and these will be applied rather than trying to build a model from scratch with a new knowledge structure.

If the system has good amount of information in knowledge structure they could make some good and interesting predictions. The author has explained by taking an example of horse who died in stable. A good knowledge structure and some good patterns is needed to come to a solution to the cause of death.

The biggest problem we face while an AI system modify itself is that of trust and those modifications are safe. This means that we need to know something about all possible modifications, but it is very hard in case of Artificial Intelligence Systems to decide whether those modifications are safe or not if no one can predict those changes.

In recent times it is still hard to decide whether these self-improvements as safe or unsafe and it is still exceedingly complicated. In fact, preventing the construction of one specific kind of modification is so complex that it will require deep understanding of those self-improvements.

PROBLEM: Generalization

Generalization is an important part of AI. Machine learns not only by analyzing failures but it notices regularities and patterns in the given data and drawing out a generalized solution or a required input from them. A IPP program was designed based on this problem which augmented store of its predictive Knowledge structures by remembering the generalities and patterns from the data.

Presently, Goal of generalization is to make AI system perform better on test data. Goal is to force neural network and other machine learning algorithms to learn useful concepts in one scenario and perform better on new ones.

PROBLEM: Curiosity

The author explains example where the story is recited to students and they are asked to ask questions after every pause they are curious about. This would result in better understanding of the topic as they would analyze every part and make hypothetical scenarios.

He further says that everyone is curious about what they see, what they hear, what they are told. This is a natural result of a system making predictions and wondering why those predictions fail. One view of curiosity is that a system is curious if it devotes processing effort to identifying gaps in its knowledge base and seeks to fill them.

So, the conclusion can be drawn that the system should be curious about the predictions it is making and seeks why these predictions happen.

Presently, curiosity AI is the simulation of human curiosity in artificial intelligence. Curiosity AI is important as it can hold and look at more data than a human is capable of, while being able to view that data in a more human-like, critical way.

PROBLEM: Creativity

This problem discusses about the creativity. The creative explanations misapply and use an indefinite pattern to make the explanation creative.

Creative explanation and understanding involves taking a pattern or knowledge structure and modifying it to fit events other than the ones for which it was originally derived.By following an inappropriate explanation pattern and making it compatible with the current problem, an explanation is derived that is creative.

Creativity, whether in the artistic, scientific or problem-solving domain, involves using something in a novel and unexpected way.

By creative reasoning the problems of retrieval and application were reducedthe main focus while creatively searching for knowledge base is to be intelligent about which “inappropriate” structures to use as machines does not possess infinite processing power.As inappropriate patterns are selected, the problem of creative reasoning can be solved.

Presently, Creativity may be the ultimate moonshot for artificial intelligence.

CONCLUSION

The problems discussed here are set of goals and the benchmarks that we might use.

The field of AI goes back to the 1940’s. This concept grew academically in the 1950’s. The next phase was in the 80’s where AI techniques were being moved from academic applications into the business world. During the 80’s and 90’s, robotics entered factories.

In 2015, we had Artificial Narrow Intelligence (ANI). Narrow AI is programmed to perform only a single task. Today our world exists in Artificial General Intelligence (AGI). It holds human level intelligence, decision making, reasoning and thinking. In the coming years, Artificial Super Intelligence (ASI) will surpasses human intelligence in all possible aspects — from creativity, to general wisdom, to problem-solving. Machines will be capable of exhibiting intelligence that we have not seen in the brightest amongst us.

In the future, AI will continue to grow and will enable new business models and functionalities. It will also play a key role in improving quality of life through allowing humans to complete tasks more efficiently and in less time. Although we cannot be sure of the exact future, it is quite evident that interacting with AI will soon become an everyday activity.