Jump to the ERE Overview page.
Jump to the Comparative Capabilities page.
Jump to the Home page.

Capabilities

Planning

ERE was designed as to be a planner, among other things. It is accomplished mainly in the Projector module which when given a strategy from the Reductor, builts a graph of possible future behaviors. The behavior constraints are used to prune the tree of unacceptable behaviors.

Prediction

Prediction is possible in ERE because it has a Causal Theory which contains a description of:

a description of the control actions that can be taken by the system is required. Without it, the system could produce no behaviors. These actions are defined by their preconditions and probability distributions of their possible effects.
exogenous events that are outside of the system's control can be optionally supplied. This information will be used primarily by the Projector to reason about possible system behaviors.
domain constraints must be provided. These constraints specify those facts which can never co-occur; these constraints are used by all system components to maintain the consistency of the world model.

Reactivity

ERE achieves reactivness through the Situated Control Rules (SCRs) that are either provided by the user or are compiled while planning. These rules are used by the Reductor to select an action for execution. When Reactor senses a change in the environment, an applicable SCRs will fire and recommend an action.

Reactivity is also achieved because the Reductor and Projector can receive sensor signals. The Reductor can notice if changes in the environment have made its problem decomposition invalid. The Projector can notice if changes in the environment have made its behavior graph invalid. The latter is very difficult and hasn't been implemented as yet.

Taskability

Their definition of taskability is that the system can accept a new goal anytime. This is possible in ERE. The Reactor can accept any goal as long as there are appropriate SCRs. If there aren't then the Projector would have to be invoked to compile SCRs. If the strategy being used the Projector at that time is inappropriate for the new goal, then the Reductor would have to be invoked to do so. This could lead to significant computation, but note that it is a conservative approach in that the computational demands propagate back only as far as they need to.

Learning

Currently, there are three types of knowledge in the system that can be acquired and refined:

Causal Theory can be refined by detecting and recovering from failures. These failures detected when there are discrepancies between the real state of the world and the projected state. These failures may be cause from either missing operator preconditions, missing operator outcomes, or missing domain constraints. An EBL technique has been implemented that resolves the first cause. Work continues on methods for overcome the other two causes.
Situation Control Rules can be acquired and refined through caching goal-satisfying behaviors synthesized by the Projector. SCRs can be formed with a vary in their degree of generality. This process is not unlike the goal regression, knowledge compilation learning technique called chunking that is used in Soar.
Problem Reduction Rules can be refined in the face of inappropriate problem-solving behavior. Recall that ERE can perform in an anytime manner: the Reactor may be forced to perform an action before the Projector has completed planning. It is possible that the action taken was inappropriate in the long run. The system would have to backtrack. The probability of backtracking can be reduced by incrementally refining the problem reduction rules such that the subproblems are more independently solvable.

Navigation/Manipulation

Since the architecture is designed for planning, scheduling, and control, it is certainly conceivable that it is applicable to robot tasks where navigation and/or manipulation is required.

Coherent Behavior

ERE is considered to behaves coherently and rationality because it has the ability to act at anytime and uses all the knowledge that it has at its disposal, whether it was synthesized or provided by the user.

Perception

ERE must be interfaceable to sensors (see Reactivity). There weren't any details about this in the reference however.

Other

ERE was designed to integrate planning, control, and scheduling. Scheduling is the ordering, in time and space, of operators in accordance with a given set of constraints. This is the only architecture we've studied that was concerned with scheduling.