Reflections on Collaboration

Humans are rationally bound, context driven, agents and try to feed this contextual drive by gaining additional perspectives. Working with other tools or agents helps to fulfill the contextual gap, but in working with automation it is easy for this act of gaining perspective to fall out of the realm of collaboration. The fundamental first law of cooperative systems of “if you do it all or I do it all, its not collaboration” is immediately broken if the agents do not work together to share their perspectives. An agent simply looking through a camera on another agent is doing all their own work; they have to use their biases to figure out the world, they need to tell the remote agent with the camera what to do in order to build their mental model of the world, it makes no difference if they were interacting with a walking camera instead of a human agent – if there is no communication between the agents there is no collaboration between the agents, and this is the immediate trap robots as agents can fall into.

When two human agents have an open line of communication they can collaborate with each other and the remote agent can embody themselves in the localized agent, gaining perspective. If that line of communication is closed, or one-way, the collaboration falls apart. The remote agent surely gains more than if they did not have access to the video feed, but they are hardly privy to the other agent’s perspective.

Just as two human agents without communication are not collaborating, an agent and a robot with a camera are not a collaborative group. The robot and human are fairly independent parts that have very limited interaction when integrated: one looks, and the other postulates, there is limited, at best, communication between the two. It was thought that if the agent controlling the robot felt sufficiently embodied in the robot they would “collaborate” better, but is it true? If the human does all the work to make their perspective match the robot’s, or if they program the robot’s perspective to match their own, they are defeating the purpose of using automated agents.

The entire goal of collaboration is to capture the diversity of benefits that mixed groups can provide, while minimizing the difficulties in communication that the same diverse groups encounter. Embodying an agent within another agent only provides the benefits of collaboration if they agents are actively communicating with each other and if the agents actually have different perspectives. Simply providing a remote camera does not bring any diversity to the situation, a camera does not have its own perspective.

The other thing collaboration brings to the party is trust – as per family 3 and Mr. Weasley’s rule: “Never trust anything that can think for itself if you can’t see where it keeps its brain.” As we talked about in class, there is no way a human agent is going to trust a robotic agent with their life unless the human agent understands what the robotic agent is doing, and why. If there is understanding, trust is much more likely, and the only way to craft understanding is through communication. When the robotic and human agents are able to communicate the human agent is able to gain insight into the robot agent’s perspective, the why behind the what, and act accordingly. This insight is clearly important, but when it replaces individual perspective collaboration breaks down again.

When the human agent is no longer be able to distinguish when their perspective gains are actually things observed, or simply suggestions made by the machine we are back at the first fundamental law. Just as when an agent starts providing solutions the resulting bias on the other agents hinders collaboration, when a human agent feels that they can take the robotic agent’s assessment as the only plausible one, collaboration has broken down because only one agent is actually doing any work.

The key difficulties in maintaining collaboration are keeping communication channels open, providing access to alternate perspectives and knowing when new, fresh, perspectives should be brought in. In automated systems making the robot’s brain clear is a sure-shot step in the right direction, providing the human agents working with them the context they need to not only collaborate, but to act.

-Brian

Cross Scale Interactions

For resilience, you need to look at more than just decomposing and isolating levels, you need to look across them. “Turn it sideways” if you will, and get away from the vertical, subordinate levels of control. Each locus needs to have some authority, which sets up supervisors as place/stakeholders in addition to just delegating authority to local agents(possibly robots!) Setting up a authorty delegation system based on the cross scale interactions makes it different than a command structure, because its not a strict hierarchy of authority, and it would be much more stable in the face of change.

-Brian

Agent (and Eliza effect)

Agents have both intention, and the ability to act on their own. The Eliza effect is when machines confuse the people interacting with them by appearing more animate and “agent-like” than they actually are – inferring intent where there is none.

-Brian

Ackoff’s Catch

Simply, the developed model of a problem is never able to produce a completely accurate model of the solution to the problem unless the model of the problem is perfect (which is never is).

-Brian

Questions

Can you actually define goal? Is a goal simply an end-state? A desired end-state? How many criteria are needed, and how many people need to agree on or share those criteria for an end-state to be considered a goal?

What are the interactions between vulnerabilities and opportunities you get for control in coordinated systems, and why do you abandon these in favor of single agent work? Different people defect at different vulnerability:opportunity ratios, so which do you use when designing an automated collaborative system?

A shorter one from class that I thought was important enough to warrant repeating: Does automation increase or reduce human skills?

What’s the difference when you are working as a team and can see each other, vs not seeing each other? As humans, we’re very good at directing attention, so we clearly lose something when we lose the ability to judge perspective, but robotic agents do not share that trait, so how does their perspective taking affect them?

Does 1 failure mean that you no longer trust the machine? Where does 1 experience of failure change trust?
When the failure is understood by the other agent they are more lenient. In some cases, if the agents
understand why the failure occurred they may even trust the robot more; but if they don’t understand why
they will almost always lose trust.

Staying on the trust issue: How does trust vary between levels of capability versus experience? I trust you because you’re good at this job vs I trust you because we’ve worked together forever.

Trust can lead to fragmentation, agent can trust an agent to do some of its tasks, but not all of them, and still collaborate with the agent. Why is it easy for this fragmentation to occur, and why is it harder to “patch up” and fix than simply creating/hiring another agent?

In the case of supervisor systems, does the supervisor need to be more expert or less than the first level controllers, and how does this dynamic relate to the generalist vs specialist problem? Is the supervisor a generalist despite previous specalist experience, or vice-versa?

And of course, the ultimate question: What are the limits on single agent activity that force us to work on a group ensemble project? What are the costs that move us from coordination to single agent activity?

-Brian

The Leftover Principle

From the last day in class, thought I’d include this for completeness.  The leftover principle states that we should just let the machine do whatever it can do, and then let people do whatever’s leftover.

- Mark

Some Reflections on Coordination Issues

Is coordination “worth it” (and for whom)?

A few years ago, Japanese bee keepers decided to introduce a European bee species into Japan in order to increase apiculture production. They had realized that these bees were allowing them to produce about three times as much honey as the Japanese bees. Unfortunately, they also discovered that one specific Japanese wasp species was sometimes decimating European hives, which didn’t happen with indigenous bees. The reason for this is a totally different way of managing a wasp intrusion into the hive, an inadequate defence mechanism on the European side and a highly performant adaptation of the Japanese bees.

Confronted with a wasp attack, the European bees reproduce the defence strategy they use in Europe, which consists in successive individual attacks. Unfortunately, Japanese wasps happen to be much bigger and stronger than European wasps (that the foreign bees are used to being confronted with – and are thus adapted to) and the individual bees attacks present absolutely no threat to these wasps that mechanically kill their inoffensive opponents one after the other until there is no more attack (what sadly means that there are no more bees). Through evolution, on the other hand, Japanese bees have developed a remarkably successful way of dealing with these dangerous intruders: when a wasp enters the hive, some alarm mechanism plays (that I don’t remember precisely), and a dense envelope composed of dozens of bees quickly forms around the aggressor. First, this is a way to prevent the wasp from moving further and to make it harmless. The bees then start violently shaking their wings while keeping the “bee ball” extremely dense around the wasp, which results in a progressive rise in temperature inside the ball due to energy expense. Within a relatively short time (I think it is in the range of a couple minutes), the temperature is hot enough so that the wasp body temperature exceeds its limits, causing its death, while the bees show a higher threshold and therefore remain alive.

To my mind, this example illustrates several key aspects of coordination and some common issues. First of all, we can clearly see that coordination is not about aggregating individual ways of dealing with the situation: it is about doing something else thanks to new possibilities. Then, to discuss the costs and benefits of these two situations, we have to consider different agent and time perspectives:

• From the bees’ perspective, which can be considered the local agent’s perspective, the coordination situation surely represents a considerable benefit, especially considering short-term perspectives. It would be interesting to analyze the effects of this strategy on a longer term basis however (for example, isn’t there any harm due to the increase in body temperature, even if it doesn’t cause immediate death?).

• From the bee keeper’s perspective, which can be considered analogous to the supervisor’s, the short term perspective also shows benefits in the coordinated response since the hive is not destroyed, allowing the honey production to go on. But on a longer term basis, is it still worth it? How is a potential hive loss balanced against the much greater honey production? Of course, if these agents were human, for example in a military environment, the perspective of their destruction would be highly considered (especially nowadays – at least I hope so). But, without trying to be cynical, for insects or other animals, or automated agents, to whom we don’t show the same moral attachment, their destruction could be envisaged.

 

Is Open Workspace always a solution?

Considering a situation when it is physically possible, is there always an advantage to working in an open workspace? Such a configuration definitely presents benefits in the coordination mechanisms that would be needed in a collaborative task because it allows common ground, observability, directability, etc.

These points have been highly discussed in class and are fundamental aspects of the collaborative projects presented parallel to this paper: see “Common Ground” by Zelik and Branlat, and “Open Workspace” by Fern and Anders. In this second project, video illustrations show how distracting working in open space can also be, especially when non-working phases of other agents are involved. Moreover, in the “Common Ground” video, work itself is causing difficulties because of the noise generated by all the potential communications, and this to such an extent that the emergency call centre had decided to have ergonomists redesign the workspace. Through their analyses of the activity, ergonomists showed how the noise was not just noise, but was also conveying important information used by the team members to coordinate and fulfil their mission.

In this working situation that demands such a level of communication and coordination, the cost of the noise due to the openness of the workspace is balanced against the benefits it generates. In other less collaborative working situations, I don’t think this is necessarily the case. Taking the example of the CSE lab, its members work to a large extent on rather individual projects: classes assignments, readings, research projects, etc. Being in an open workspace has several benefits, even for these tasks: other people can be a useful resource for helping to solve a problem or recommending relevant readings. In this last example, common ground, especially some knowledge of what each person is doing, highly favours such interactions. However, there are periods of time when a need to be highly concentrated on one individual task arrives, and the openness of the workspace thus becomes very costly and counterproductive. In such a situation, calm isolation can be hard to achieve (staying home could be my personal strategy, although it causes other problems, presenting other sources of distraction).

 

How much should an agent know about the goals?

Participating in the “Convoy” project as a “guinea pig” made me conscious of an interesting aspect of coordination: I was the last person of the convoy, and my task was to “follow the previous person”. However, and contrary to other agents of the convoy, I had no further information about a higher level, collaborative goal. I therefore decided to scrupulously follow this rather simple task by keeping an eye on the person in front of me. This presented no particular difficulty and I can say that, from my perspective only, I completely achieved my mission. However, because of the conditions of the experiment and the decision made by the agent in front of me (who was dealing with other issues), it was simply not possible for me, without a broader understanding of the situation, to help manage the error that had been made. So from the system’s perspective, the collaboration was a failure.

In this example, the cost of coordination seen as allowing agents to have an understanding of the whole situation is a complexification of the agents’ tasks: I would have had to solve a more complex problem. The benefit is however a greater resilience for the system: my decisions were totally dependant from the previous agent, definitely creating some brittleness in this incompletely collaborative system.

- Matthieu

Another list of big questions…

Is detection-only automation more collaborative than full decision automation?

Is it safer to leave the decisions to human agents?

How can an agent be observable enough by all the concerned stakeholders without giving too much information?

Can an automated system be considered expert?

Should an automated system be implemented to be creative? Is it possible?

Is human interaction the right model for designing coordination between various types of agents?

- Matthieu

Interruptibility

Is judging if an agent is able to and at a point to receive a new message. This implies that the interrupting party must have adequate detail to know the state of the other’s current activity. (JCS book)

- Shilo

Sheridan’s levels of automation

Sheridan’s levels of automation treats the question of automation as a simple, symmetrical one-dimensional issue, with a point being gained for computers and lost for humans, or vice versa. At the big end of the scale, humans are not involved. This ignores the involvement of humans as collaborators and stakeholders.

- Shilo