HOMER

A diagram of this architecture

Philosophy

The underlying philosophy behind the creation of Homer, was integration. Vere and Bickmore believe that the research into separate components (planning, learning, etc.) has progressed sufficiently to allow the creating of an interesting integrated architecture and agent.

The architecture on which Homer exists is a modular architecture. It consists of a memory, a planner, a natural language processor, some monitor processes, and a plan executor. This modular design is due to their ambition to integrate several research into several AI components into a single working agent. In the future, they plan in include an inductive learning module. There is no true central control module, although one might argue that the planner fulfills most of this function.

The memory can be divided into two sections, general knowledge, including world knowledge and lexical knowledge, and episodic knowledge. The planner is interesting is that it is a temporal planner. This planner as well as its associated reasoner are derived from DEVISER V. The natural language processor accepts input from a human (via keyboard) parses it, and outputs results to a screen using a sentence generator. As mentioned, lexical knowledge is contained in the general memory. The plan executor works closely with the monitor process to attempt successful completion of the plan. The monitor, or reflexive, processes provide feedback to the plan executor in case replanning is necessary.

Agent Properties :

One of the most interesting properties of this architecture is its temporal extent. This is due to its episodic memory and its temporal planner (discussed below).

The knowledge base in Homer is global . Any of the modules can access it, although only the input modules, such as the text interpreter and world sensors, and the reflective processes can modify it. Knowledge is represented in a frame-like manner. Knowledge consistency is handled by simply stating that the most recently learned fact is the true one.

Capabilities :

The temporal planner does not just plan actions to follow one another, it plans the actions to happen at a certain time, or within a certain time frame. The planner is capable of replanning. For each plan, there are a set of protection conditions. Whenever a fact is terminated from the knowledge base, the protection conditions are checked. If they are no longer valid, replanning is necessary. The planner contains a separate episodic memory of its own for backtracking if necessary.

The natural language processor serves a dual purpose. It accepts text input from a user which gives commands to Homer. This provides the system with taskability . The processor also provides feedback to the user. Possibly this capability could also be used to facilitate cooperation between agents with similar NLP abilities.

Homer has the capability to sense its environment. It also has the ability to perform limited actions within its environment, such as moving, grabbing objects, and releasing objects. These abilities, combined with its memory of the world allow it to navigate within the world.

Homer is also capable of learning facts. In this it is limited to knowledge it can learn through its sensors, natural language processor, and monitor processes. It does not have any learning techniques other than these.

Environment and Agent Body :

Homer is designed to function in a limited dynamic environment. What makes this environment limited is that it is not the real world, but rather a simulated environment. This simulated world gets rid of many of the real world problems. There are only a limited number of objects, and they act in very limited ways. Actions and perceptions are assumed to be perfect . Additionally, classification of objects is assumed to be perfect. Supposedly such classification is a solved problem for underwater agents. Further, self-location is assumed to be perfect. Also, since this is a simulation, the actions performed by the agent are very limited: move, pick up, and drop. There is only a single speed of movement - currents do not exist.

Even existing in such a limited world, where only a very limited amount of knowledge can be obtained, Homer suffers from slowdown if the episodic memory grows too large. Plan generation is also very slow.

Issues :

As previously mentioned, the natural language interface provides Homer with taskability . Homer has two main efficiency problems. The first is the planner, which is a bottleneck. The second is the lack of scalability in the episodic memory.

Other Architectures