Jump to the RAPLH-MEA Overview page.
Jump to the Comparative Capabilities page.
Jump to the Home page.
Capabilities
Planning
This architecture implements planning in the traditional sense. An
explicit plan is formed and executed. However, it also does replanning
if the utility of replanning is greater than the utility of sticking
to the current plan. It does this metalevel thinking by treating the
action of replanning just as any other action. The architecture is
such that it allows it to pick, for execution, the next action which
has the greatest utility.
Prediction
There are various types of knowledge kept in this architecture. Some
of them are: knowledge about possible next states that will follow a
given action, and knowledge about the utility of the next state, etc.
These types of knowledge about the world are, in fact, rules that help
the architecture predict how the environment will behave; either by
itself or as a result of the agent's actions.
Meta-Reasoning
RALPH-MEA was developed, from the start, as a meta-reasoning
architecture. It borrows heavily on Russell's work on meta-reasoning
for limited rationality. Specifically, it can decide when it is wise to
do replanning (i.e. it decides when to think more rather than execute
a physical action). This allows it to make smart, on-time decisions
about when is wise to abort the original plan and create a new one to
handle some new and unexpected situation.
Reactivity
One of the EA is the Condition-Action module. This module is a
reactive module in the sense that it decides what action to take based
simply on the current conditions of the environment. However, wheter
or not this action is taken is still decided by the metalevel module.
Suppossedly the metalevel module will have enough knowledge to
determine when a quick action (i.e. one from the CA module) is the way
to go. This, however, was not make clear.
Taskability
The architecture was designed to handle tasks. These task, however,
can be interrupted or preempted by other more pressing emergencies.
The submarine's main task is to gather information about possible
mines and to map their locations. It will do this as long as it does
not interfere with its main goals of staying in functioning condition
and avoiding enemy radar.
Learning
The authors explained how knowledge from one of the EAs could be
compiled into knowledge for another. This might be viewed as a type of
learning (akin to caching). However, they did not mention how the
agent would go about learning the tonnage of probabilistic domain-knowledge
that it needs to function.
NLP
It does not implement NLP.
Interruptability
The architecture is highly interruptable. It monitors the enviroment
constantly and any new incident is taken into account.
Navigation/Manipulation
It does not implement these since the agent ran in a simulated world.
Coherent Behavior
It shows a great deal of coherent behavior. Not only is the agent
goal-driven but it also knows know to abandon these goals (or plans)
if a conflicting problem arises. For example, when gathering data, the
submarine will keep doing its job unless one of its batteries dies. If
so, it will cut some part of its plan and return home earlier.
Perception
The submarine works in a simulated environment but it does implement a
simulated radar and sonar sensors, plus some other instrument readings.
Other
This architecture also possess the capability of limited rationality.
That is, it is capable of making smart decsions about when to act and
when to continue thinking. This capability is implemented using
meta-reasoning and decision theory.