CSCE475/875 Multiagent Sysetms

Handout 9: Collaborative Topic Summary Assignment 2: Q & A

September 16, 2009

>> Several questions were selected from Assignment 1. Here is my response to these questions.

Q1. Is it feasible to have nested agent societies? This is easily conceptualized in nature where we have cells that make up tissues, which in turn form organs, which make up systems, then beings, and depending on the species, societies themselves.  Is there a practical utilization, other than to perhaps model nature itself? If the practice of nested agent societies does occur, how is information typically passed from the subagents to the super agent so that information can be shared with the super agent’s co-agents …?

Response: This is a loaded question. First, what do we mean by “nested agent societies”? A group of agents form a society. And that society belongs to another society, and so forth. Or, there are super agents, under each super agent, there are several sub-agents. Under each sub-agent, there are several sub-sub-agents. Which one? In my opinion, nested agent societies mean societies within societies. So, the example clarification provided above “ … easily conceptualized in nature … cells that make up tissues …” is quite accurate. Now, is there a practical utilization? Look around us. Do we see such nested societies configurations? Yes. We do. In a university system, for example, we have staff, faculty, students: your three basic societies. They all belong to the big umbrella society of the “university people”. Under each of these societies, you can have sub-societies. Why do we have this kind of distinction, say, on the university blackboard site, or phone directory, or regular department websites? Better organization. And thus, yes, there is a practical utilization. And for the question on information passing. Once again, look around us. How do we pass information up and down the “command chain”? Say a student informs his/her instructor about a problem P. The instructor considers and decides whether to share this problem P with his/her colleagues. And so forth.

Q2. What is an effective way to measure the coherence of a multiagent system?

Response: Suppose that we consider coherence as desirable emergent behaviors. For example, stability, optimality, etc. Stability means the agents in the system converge to a steady state when handling events observed in the environment. Optimality means the system as a whole finds the best solution for each given task. And so forth. Now, how do we measure stability? When agents make decisions and commit to them without flip-flopping or switching to other options with no end in sight. So an effective way would be to measure the number of times agents change their decisions after a task is introduced into the system. And so forth.

Q3. How can we measure the group competence ability of a particular agent?

Response: First, what is group competence? It is a measure of the overall group in terms of its performance as a group. So, now, what do we mean by “group competence ability”? How well agents can work together in terms of teamwork skills? If that is what is referred to here, then teamwork can be measured in the statistics that we use in class for the ClassroomWiki assignment, for example. Number of messages, timely response, number of collaborative actions, joint decisions, etc.

Q4. Would intelligent agents in a large society of other autonomous intelligent agents begin to form groups on their own as do humans do? Maybe each believing a certain goal is better performed in a certain way?

Response: Yes. In MAS research, agents do form coalitions, or neighborhoods. However, it is not always because they each believe a certain goal is better performed in a certain way. Usually, these agents form their coalitions to solve problems of concern to them. Not all agents are needed to solve a problem. Usually, a subset is sufficient. And thus a group can be formed by the agents on their own. (My research work is in coalition formation. So, if you are interested, please come by and we can have a long chat.)

Q5. Are agents capable of performing altruistic actions (neither advantageous nor disadvantageous) when there are no other immediately available agents that could work on a task?

Response: This is a tricky question. I brought this up in class. Philosophically, can a human being be altruistic? Suppose a person sacrifice himself or herself to help people. For other people, they might view the person altruistic. However, the person might say, “I feel good when I help people.” Thus, if you measure this person’s utility gain when helping people, it might be really high especially in the dimension of “happiness”, for example. In that case, we could say that this person is still selfish, self-interested. But this person’s view what is important is the happiness that he or she experiences when helping people.

Now, we come back to the question at hand. First, would we as MAS designers design such a system? A system where it is possible that no available agents can work to solve a task or choose to solve a task because of low utility? Probably not. Even if it occurred, one would probably want to revise the design to prevent that from happening. Unless programmed to react blindly, agents should have motivations in what they decide to do or not to do. Also, if an agent is capable of carrying out a task, but somehow does not bid for it (for example), and only does it because other agents cannot work on it, then do we say this agent is altruistic or self-interested? Think about this. Furthermore, if somehow this scenario could occur in our system, does that mean we have designed the local decision making process of each agent correctly?

Q6. Utilizing the Contract Net protocol, would a previously occupied agent respond to an open task or would it wait to bid until it has finished its own task (since it would still receive the query message despite being busy)?

Response: It depends. If the design involves prioritized tasks, then it is possible that an agent might want to bid for a more important task while it is still working on its own task. Thus, think about the problem characteristics.

Q7. The interaction protocols covered (Negotiation, Contract Net, Blackboard System) seem to be based upon abstractions of "human ways" of interacting. Are there any protocols that rely on agents behaving in ways that humans wouldn't?

Response: There are many interaction protocols that are based on animals: bees, ants, swarms, etc. In those protocols, there are behaviors such as ants leaving pheromones to conduct indirect communication instead of direct communication, which could help solve the travelling salesman problem quite effectively.

Q8. Are there any existing multiagent systems (outside of game scenarios) where agent deception is condoned or encouraged? If so, are exceptions made for human agents in such a system?

Response: There are two views to this. In general, if you are the sole MAS designer for your MAS, then it would not make sense for you to want to implement “deception” capability into each of your agents. That’s the first view. Now, if you design a group of agents or one agent to run in a multiagent system that is open – designed by many other people, then perhaps agent deception can be encouraged or condoned. For example, in cybersecurity research, harmful agents can be designed to attack the Internet and see how resilient the Internet can be. Human agents are different though … there are legal laws and social rules governing how humans should behave in a group or in a system.

Q9. Instead of designing agents that construct an internal representative of other agents, can a communication protocol be designed that allows agents to refer to the other agents?  In such a system agents would always have access to a perfect model of any agent, which helps to predict future behaviors.  This system would involve agents performing additional action simulations, which would delay their time to complete other tasks, but eliminates the space requirement for explicit agent models.  Are there disadvantages to this approach?

Response: The short answer is yes. Especially when communication is cheap and any delay would not hurt the quality of the solution. But, let’s look at two aspects. First, suppose that agent A1 needs to make a decision that requires it to know of certain aspects of A2. Since it does not know of A2, it sends a message to A2 to ask for that information. A2 may or may not comply, depending on its motivation and utility gain. And then A2 proceeds accordingly. In this case, it is perfectly fine. Now, what if A2 is busy, what if A1 needs to do this all the time and has to contact hundreds of agents, etc.? Without profiling, even if it is inaccurate or incomplete, A1 would have to at a given time window, try to collect all those information in order to make a decision. With profiling, however, A1 would be able come up with the best decision it could make given what it knows. That’s an advantage.

Now another aspect: suppose that agent A1 needs to make a decision that requires it to know of certain aspects of A2. Instead of soliciting information from A2, it sends the “task” over to A2, passing the responsibility to A2 with A1’s own information. Would this work? First, why would A2 want to take on the task? What if it is simply something that A2 just cannot do? Now, here is then a potential spam problem.

Q10. It seems plausible that the essence of intelligence itself is unintelligent—that it is the side effect of a completely mechanical process which is repeatedly applied to environment which includes both the system executing the process, and the process itself insofar as the that environment is perceivable by the system (which is itself a component of the environment).  This allows the phenomenon of intelligence to be described in similar way as phenomena such as weather, since it is both a component of the environment and coincidental to particular environmental conditions (where the coincidental conditions that might allow intelligence to take place are analogous to those that allow particular weather to occur).  What might such intelligence procedures look like?

Response: In class, I mentioned that in AI, some argued that human intelligence is emergent of something very simple: individual neurons and connections between neurons. In that case, the “intelligence procedure” consists of the “signal threshold” formed in each neuron, and how each synapse is fired. And these could be argued as the product of positive and negative reinforcement over time; and some could also be argued as evolution over time and humans are born equipped with. I do not quite agree with your analogy with the weather in terms of “both a component of the environment and coincidental to particular environmental conditions.” Weather is an abstract concept, a “snapshot” of our environment. There are different components such as cloudiness, temperature, winds, precipitation, etc. that allow us to better describe weather. But that is different from your assertion.

Q11. In speech act theory, human natural language is viewed as having three aspects: (1) locution, or the physical noise produced by a speaker, (2) illocution, or the intended meaning of a spoken noise, and (3) perlocution, or the actions caused by a spoken noise.  That is, a particular phenomenon which occurs in an environment for which a meaning is understood can be referred to in many ways.  Can an agent be designed (e.g., using an assumed "intelligence procedure") which identifies artifacts in the inherently meaningless noise which it happens to be able to perceive, such that it can produce meaning, and ultimately construct a theory for the true environment where it exists?  This is somewhat fundamental question which does not necessarily assume that a perceivable environment exists.  For example, assume some apparently sensor-less entity situated in the natural world is an agent which executes an intelligence procedure.  Despite the possibility of intelligence in the entity due to the presence of an intelligence procedure, it cannot achieve intelligence with respect to the natural environment since it cannot possibly perceive that environment (furthermore, perhaps what we perceive as nature is a subset of the environment that truly exists—in that case, the entity may be intelligent with respect to another environment which humans cannot perceive).

Response: Yes. (If you are interested, read the novel “Contact”. ) There is what some AI scientists have come to agree to be the true form of intelligence: learning by discovery. Such a system is capable of mining the data it observes and detecting meaningful patterns. Researchers in data mining and knowledge discovery are at the forefront of this line of work.