Refining the definition of a state of understanding

[chcorbato]

from D2.2 understanding candidate:

If the state of understanding for a cognitive agent is the state of being in possession of a model that can be used to make sound inferences \cite{D1.3}

I think we need to align on the formalization the definition of understanding state, since they are the goal for this understanding logic.

@alex-quin-gabriel can you formalize and share here your definition here? If you have in TPTP, maybe share the code and explain it for those that do not know the syntax.

Also, the motivation is to make "sound inferences". But there are many possible inferences. It may help us reducing the socpe, by identifying the specific types of inferences done in an autonomous cognitive robot control system.
what do you think @rsanz @alex-quin-gabriel

[chcorbato]

I think we should use concrete needs to drive this definition, for example, taking the Wind estimation example, maybe @GPatiA2 could help us define how the need for that wind estimation arises, who generates that understanding goal?

[GPatiA2]

The need for estimating the wind comes from the fact that operating a drone in a harsh environment with lots of wind is unsafe since it may crash due to a wind gust and may not be able to complete the mission autonomously. In general, the weather is decisive in whether a drone can be operated, and the criteria to meeting that constraint or not lies in the operator.

[rsanz]

The question is not that the engine decides to make sound inferences instead of doing other, non sound inferences. It is a question of 1) the model fidelity to the target system and 2) that the inference procedures are sound (derive true from true). Deciding what inferences to do depends on the way of operation of the engine (e.g. a problem solving strategy directed by goals vs directed by facts).

[alex-quin-gabriel]

So the 'why' of the wind model arises from

• the agent's values: 'safety', 'effectivity'
• via its self-understanding: 'I cannot fly safely in high/gusty wind'
• via its mission-understanding: 'I need to inspect the plant to be effective' -> 'I need to fly'

[chcorbato]

Nice analysis of systems requirements about understanding, I'll update all this info in D2.2 after extracting more modelling and design decisions in this thread. :-)

This starts looking like a nice scenario description:

wind model arises from:
- the agent's values: 'safety', 'effectivity'
- via its self-understanding: 'I cannot fly safely in high/gusty wind'
- via its mission-understanding: 'I need to inspect the plant to be effective' -> 'I need to fly'

So, in terms of proposing a candidate solution to this scenario:

@rsanz , which role (functions) do the Understanding core elements play?

- model: synthetiser
- model integrator
- engine activator

Do they cover all functional needs in the scenario?

Coming back to the external (to the understanding system) actors:

• is the wind estimation need required through the entire mission for all missions in the testbed? @GPatiA2
  if so, there is no need to dynamically generate the pipeline for wind estimation, right?
  another case could be that, even if the need to have an estimate and wind estimation piepelines are fixed, something changes during operation that requires to change the estimation pipeline (e.g. the estimated wind force gets closer to operating conditions, but the confidence in the estimation is low or its precision known to be poor, so there is a need for a more accuate estimation than the existing one).

Which elements and which functions would playe in such scenario?

[alex-quin-gabriel]

I guess you could vary the frequency of wind estimation depending on how bad the wind is....but there is also the problem of wind-shadows.
a drone sitting on the ground or in a truck could be shielded from the wind by buildings, vehicles, solar panels, surrounding trees....and encounter different wind conditions at different points in the mission, depending on how the environment interacts with the wind.

a drone in a swarm might favour the wind estimations of other drones in its direction of travel, or even older (up to a limit) estimations if those are closer to the path it wants to take.

this probably gets closer to scenarios relevant in the testbed...but you could also imagine a drone that has tasks inside and outside, and can forego wind estimations while being inside..

[GPatiA2]

another case could be that, even if the need to have an estimate and wind estimation piepelines are fixed, something changes during operation that requires to change the estimation pipeline (e.g. the estimated wind force gets closer to operating conditions, but the confidence in the estimation is low or its precision known to be poor, so there is a need for a more accuate estimation than the existing one).

I'd say that @chcorbato 's is a pretty good scenario to validate certain functions of the architecture.

is the wind estimation need required through the entire mission for all missions in the testbed? @GPatiA2

Wind estimation is indeed very important for an ideal real autonomous drone mission. However, for us it'd be unfeasible to go and run the mission in unfavorable wind conditions since we would not be able to test the rest of the system. I guess that, during real test time, we would only need it in a situation where the objective is leveraging collective awareness in the system to handle multiple wind measurements. For example, during the emergency landing scenario which we want to use to validate collective awareness during the projection and evaluation stages, having wind estimation is not very useful. I hope this answers your question :D

[alex-quin-gabriel]

I can see a situation, where changes to the wind estimation pipeline are required: while a drone is sitting on the ground, prior to start.

in this case, it cannot use its own movement to infer wind, but must rely on other sources. (other drones, the weather report, some external sensor,...).