Expand single agent builders to complete complex tasks

[lucweytingh]

For comparing multi-agent building, single agents need to be able to create constructions on their own. Building this could form a basis for multi-agent construction.

[gsileno]

Sorry, I don't get the "comparison" meaning here. Can you elaborate a bit more?

You're aiming to perform multi-agent building. Supposing the building design is given, it is a matter of planning (decomposition of tasks) and of scheduling/configuration (allocation of time, resources and work). Decomposition of goals can be done across agents and across time. If there is only one agent (with no possibility of coordinating with other agents), this will be constrained to a decomposition across time.

If there are several agents one can consider all kind of configurations in the spectrum from strictly centralized to strictly decentralized coordination. Let's suppose time is secondary with respect to agent availability, the problem is then to decide which agents does what. In centralized coordination, one agent will give orders to all the others. In strictly centralized, the orders are about primitive operations. In centralized/decentralized, they're about rather lower-level/higher-level actions. Each agent can decide/select by itself how to obtain this goals. But the agent might also ask to someone else to perform it at its place. In purely decentralized, the agents comes up by its own of what is the most pertinent goal to do, delegating to others implicitly (e.g. by signaling what they are doing) or explicitly (by requiring others to do something) what to do.

On Tue, May 26, 2020 at 3:27 PM Luc Weytingh notifications@github.com wrote:

For comparing multi-agent building, single agents need to be able to create constructions on their own. Building this could form a basis for multi-agent construction.

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/DanielPerezJensen/mapc-uva/issues/41, or unsubscribe https://github.com/notifications/unsubscribe-auth/AAJFMVGQRH5PI47P3KU5ZH3RTO7TVANCNFSM4NKJ277Q .

[lucweytingh]

To evaluate the efficiency, performance, and scalability of multi-agent construction, we need to be able to compare it to a baseline. In this case, single agent construction could form a baseline to compare the multi-agent construction to.

We have not yet decided about the multi-agent configuration. Dorian has built a way for every agent to access the beliefs of every other agent; intention access could work the same. In this case, a decentralised configuration could work. But I can also definitely see the value of centralised solution, so it is not definite.

Would you say you have a configuration preference? If yes, why?

[gsileno]

To evaluate the efficiency, performance, and scalability of multi-agent construction [w.r.t. single agent baseline]

this is ok!

Would you say you have a configuration preference? If yes, why?

I don't have any preference, as centralized vs decentralized respond differently w.r.t. context. when more resources (e.g. of knowledge and control) are available centralized solutions are plausibly faster with less resources decentralized are more resilient (and offer an unexpectedness factor that centralized do not have).

every agent to access the beliefs of every other agent

well this is not really a typical case of decentralized agency -- plausibly in this case it's not a problem, but in general I would expect that you create a huge overhead in the belief search. in any case, having direct access to beliefs and intentions is a bit like "bypassing" communication, a sort of telepathy.

the centralized vs decentralized question lies where intent is generated. e.g. an agent might be a defender and internalize all goals/plans for defense. or might be a "mutable" role, depending on order received by some other agent.

On Tue, May 26, 2020 at 4:36 PM Luc Weytingh notifications@github.com wrote:

To evaluate the efficiency, performance, and scalability of multi-agent construction, we need to be able to compare it to a baseline. In this case, single agent construction could form a baseline to compare the multi-agent construction to.

We have not yet decided about the multi-agent configuration. Dorian has built a way for every agent to access the beliefs of every other agent; intention access could work the same. In this case, a decentralised configuration could work. But I can also definitely see the value of centralised solution, so it is not definite.

Would you say you have a configuration preference? If yes, why?

— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/DanielPerezJensen/mapc-uva/issues/41#issuecomment-634065374, or unsubscribe https://github.com/notifications/unsubscribe-auth/AAJFMVGUZQERPWVKT3Y3M53RTPHWRANCNFSM4NKJ277Q .