CSCE 496/896

Topic Summary Assignment 8:

Formal Methods in DAI: Logic

Questions and Answers

November 13, 2002

First, from your topic summaries, all did not mention how beliefs, desires, goals, intentions, commitments, and know how could be used in an agent. I gave an example in class about this. It is rather intuitive and we as programmers have actually implemented these “cognitive primitives” before. For example, how do you collect and update your beliefs? How do you obtain desires from beliefs? How do you find goals from subsets of consistent desires? How do you choose which goals to pursue—your intentions? How do you execute an intention? For how long? These are what we have done before in our normal, routine programming, problem solving experience. They can actually be expressed using the BDI model.

As for the Stupid Question, what about this: If a logic only allows you to reason forward, and not backward, how useful is that logic? If that is the case, hopefully, the logic will allow me to go forward and move to a previous state that I was in—that way, I will be able to correct my decisions the second time around. This loop will be critical—a loop that moves from a failed state to a previous state before one enters the loop. This is analogical to driving a car that only goes forward. To ensure that a mistake can be corrected, then hopefully, there is a path that allows me to drive the car from a wrong destination (dead-end) to the place right before I enter that dead-end street. So, we see that it is important for a logic to provide us with that loop; indeed, some researchers have created approximated logic to create that loop, or to facilitate backward reasoning.

Q1: Why is there not a thorough logical account of coordination techniques? Is this just a new area of research or is it very difficult?

A1: It is very difficult. Recall that in Chapter 5 when we discussed Distributed Rational Decision Making? The assumptions and axioms have to be carefully defined just to handle very simple scenarios. For example, an agent knows exactly what other agents want and the environment is deterministic. When you have incomplete information and dynamic environments, then you run into coordination problems that are difficult to address. In one way, that is the beauty of multiagent systems. That is why we want to solve the problem with agents in the first place. It is because of the problem characteristics that cannot be completely formalized.

However, I personally believe that a general-enough coordination logic will be available in the future.

Q2: Can real life agents, under controlled circumstances, be logically omniscient?

A2: First, what are real life agents? Humans? Suppose we are talking about humans. Can a person, under controlled circumstances, be logically omniscient? But other people may not be logical, or rational. To be logically omniscient, then this person would have to be able to model logically how other people behave.

Second, what are controlled circumstances? Dynamic, noisy, uncertain, incomplete, open worlds? If the world is one of the above, unless the agent has a very good model of the noise, for example, in the environment, it cannot be logically omniscient.

If you mean software agents applied to real problems, and assumed to be working under controlled circumstances, then yes, such agents can be logically omniscient.

To be logically omniscient, it is not enough for an agent to be logical in its actions, but it also has to be logical in its reasoning of others’ (possibly illogical) actions.

A logical person, for example, can make use of another person’s emotional states logically.

Q3: In the discussion about the Israeli/Palestinian conflict you mentioned that there exist many different complex and/or complex utilities involved in the problem. How can formal methods be used to model a problem like this? What can currently be modeled effectively and what cannot be represented? In general, what do you see as the greatest shortcoming of formal methods?

A3: There are three important questions here. I will respond to them one by one.

First, how can formal methods be used to model a problem like this? Formal methods can be used to determine (1) the prescriptive actions – what are the rational actions, what should be done, and (2) the descriptive actions – what actual actions are carried out. We can use formal methods to simulate negotiations—to evaluate the utility values of all possible criteria and all possible combinations of deals. But it requires a comprehensive domain model to first outline all relevant parts. Second, it requires a very very rich knowledge model on negotiations—including the modeling of negotiators and respective emotions and pressures and stresses. Third, it requires a very vast and complex model of all parties directly involved and parties indirectly involved. Fourth, it requires the modeling of current events, past events, and future events. You can transfer these into beliefs and desires. You can then form consistent subsets of desires as goals. For each goal, you can attach a utility value based on your model. And then you can choose the best goal as your intention. And so on.

Second, what can currently be modeled effectively and what cannot be represented? I would say emotions and irrationality. If another person is irrational, how should I negotiate rationally taking in account of that person’s irrationality? How do I profile that person’s emotions? Currently, almost anything can be modeled effectively, but not everything can be modeled effectively concurrently in the same system. That is the problem with computer and network technology. In the future, we will get there. We will have enough computational power to model every aspect of a conflict as complex as the Israeli/Palestinian conflict effectively.

Third, what do you see as the greatest shortcoming of formal methods? This question is a good debate question as many people have different answers. In my opinion, the greatest shortcoming of formal methods is its assumptions of the world—complete information, noiseless world, without time constraints, for example. For real-world applications, those are some assumptions that are not realistic. Logic is powerful. But in order to harness that power, most applied AI people have to do a lot of tinkering to make logic work.

Q4: In terms of mutual belief, in a multiagent system, is this mutual belief prescribed, or can it be cultivated along with the agents’ interactions?

A4: Both. A mutual belief can be prescribed, or can be learned. The prescription is easy and straightforward. So, here I will talk about the learning one. Agents can learn from failures in their decisions and learn how to cooperate. Through that cooperation, each agent forms a certain model of its counterpart. It is through that model an agent can start utilizing mutual beliefs.

Q5: How to present and identify those logic symbols in agent systems?

A5: This is actually not only confined to just agent systems. Most AI systems have to deal with the presentation issue of logic symbols. It is very easy. Suppose you want to represent “AND”. Then given a AND b, you can build a method that returns true when a and b are both true. To identify them, you have a parser that looks for connectives between variables. For example, you may have something like this: (a and (b or c)). Simply go through the statement from left to right.

Q6: In regards to the beginning of the chapter, is a simplification of the formal methods always a bad thing?

A6: Not really. This is a good philosophical question. Since I am an engineer by training, my viewpoint is different from most AI computer scientist. To me, no, it is not always a bad thing. But I digress. Let us ask the question: why do we use formal methods in a MAS in the first place? What are the advantages of having formal methods? Strong logical foundations, high predictability, derivation mechanism, reasoning process. So, when we simplify the formal methods, do not lose sight of these advantages. If your simplification eliminates these advantages, then, yes, the simplification now is a bad thing.

Q7: In what kind of environment is a specific kind of logic implemented?

A7: This is a very tough question. In general, most applications use propositional logic and predicate logic. That is the basic. If you want to model human obligations, for example, to factor in that part of human behavior into the decision making process, you add Deontic logic. If you want to consider time, then you add Temporal logic. If you want to have good planning, and good analyses of choices, then you may add the BDI logic. Usually, for a simplistic system, propositional and predicate logics should do. For a complex system which may have conflicting states, you should try to use the BDI logic.

Q8: Where [do] the original belief and desires comes from? Can we add new belief and desire along the way?

A8: The original states of belief and desire come from the application- and domain-specific initial state values. For example, Bel(I have $10) for our game day. That means I believe I have $10 to begin with. Yes, we can add, delete, modify new beliefs and desires along the way. That is the key for the BDI logic. It allows an agent to move from one cognitive state to another, to choose to do a particular task at time T1 and another at time T2, to make a decision at time T3 and another time T4. And the BDI logic provides a consistent way of handling all old and new beliefs and desires.

Q9: Can an agent decommit in a society? Should we impose penalty or implement moral issue in agents?

A9: Yes, we have talked about this in Chapters 5 and 6. Agents can be allowed to decommit given that they decommit responsibly. That is, when an agent decommits, it (1) knows that the decommitment is the best thing to do, (2) lets its partners know about the decommitment, and (3) does not hurt the overall society more than not decommitting. Yes, we have also discussed in class before—a MAS can impose penalty on agents for decommitting to encourage agents to learn when to decommit and when not to decommit. On the other hand, we can also implement an agent that self-imposes penalty for each decommitment that it carries out. That has to do with moral issues then.

Q10: How to implement know-how when intentions are dynamic?

A10: I think you probably have a misconception about what know-how is. Given an intention, know-how tells the agent what to do to carry out that intention. For example, if I have an intention “I want to cook some noodles,” then I have the know-how “First chop up some vegetables, second boil some noodles, prepare seasoning, etc.” So, even if intentions are dynamic and they are, that does not affect the know-how. Usually, we have a mapping between an intention and a set of know-hows. So, when an intention becomes true, then the agent can immediately obtain the know-how to execute that intention.

Q11: What are the differences among basic logic types, logics in cognitive primitives, logics in coordination and logics in communications?

A11: In terms of syntax, there are not too many differences. Be clear about it. Logic is a reasoning tool. We can use this tool in moving from a cognitive state to another state, to plan and execute the coordination among agents, to analyze and parse communicated messages among agents, and so on. Basically, the tool remains much the same; only the applications are different. So, do not confuse the basic logic types such as propositional logic, predicate logic, and BDI logic, with the applications of these logic theories.