The ATLANTIS Architecture

A diagram of this architecture .

Philosophy

The design of the ATLANTIS architecture was based on the observation that there are differing levels of activity in the environment. These levels require different mechanisms for dealing with them, depending on what is important at what level. At some levels, planning might be important, and at others, a quick reaction time. The other main motivation was for an agent operating on this architecture to be able to operate in a continuously dynamic world, complete with unpredictability and imperfections.

Architecture :

Gat's ATLANTIS architecture is an instance of a layered architecture. There are three main layers: control, sequencing, and deliberation. The control layer is responsible for the low level control of the sensors and effectors. It contains no state, and is very similar to a subsumption architecture . The control layer receives its commands from the sequencing layer. The sequencing layer is the main control center for the agent in this architecture. It tells the control layer to initiate and terminate primitive behaviors. The sequencing layer receives information from the control layer via monitoring channels. The three layers operate asynchronously.

Agent Properties :

Since the lowest layer is so simple, it operates very quickly , and thus, so can the agent when it is operating solely within this layer. This layer also has a low use of computation resources . The monitoring channels are used to detect failures . It is much easier to design algorithms that fail cognizantly, than ones that never fail. When a failure is detected, actions can be taken to recover from it.

Most of the internal state of the this architecture is contained in the deliberative layer. Among other things, it contains a symbolic world model. However, this knowledge is only available to the deliberative layer. It is not global . As with most architectures who do not maintain global knowledge, its knowledge is non-uniform . Each layer has its own way of representing knowledge.

Capabilities :

By allocating resources to the deliberative layer through the task queue, a plan can be produced. The steps for fulfilling this plan are then put on the queue, and the control layer is monitored as it attempts to carry out those tasks. In addition to the replanning that can be done in the case of failures, the architectures tasks are interruptible. This allow for a hierarchy of importance levels , which is important when you are approached by a tiger while cooking in the kitchen. There are can be cleanup problems involved with interruptibility. ATLANTIS supports cleanup, and also allows cleanup to be ignored or interrupted if an event or observation is sufficiently important.

The learning that can be done by this architecture is somewhat limited. The control and sequencing layer have no learning mechanism. All the learning takes place in the deliberative layer. So new knowledge will have a most noticable effect in the planning and prediction capabilities and in the world model. Despite the fact that learning takes place in the deliberative layer, the learning itself is reflexive .

Environment and Agent Body :

This architecture is designed to operate on an agent in a very dynamic world . The different layers allow the agent to react in multiple ways. This allow it to function in an unpredictable environment with possibly hostile states. This is important, because in the real world, there is often only a limited time to react. For example, assume this architecture is running a robot cockroach. This cockroach is heading across the room towards some food. Suddenly, a human tries to step on it. The cockroach must have the ability to act quickly in order to survive, while still retaining the ability to replan its approach to the food, perhaps scurrying around the outside edges of the room.

Issues :

One of the main issues addressed by this architecture was how to incorporate slow deliberative computations without running into a bottleneck problem. This was the main purpose behind having a href="/cogarch2/arch/layered.arch"> layered architecture . The fast layers would handle the low level actions and emergency situations. They are not firmly tied to the relatively slow deliberative.


Other Architectures

Back to the Title Page