After defining some basic concepts related to agent architecture it is important that we understand how to provide agents with the means to complete tasks.
We want agents to be able to complete the tasks we specify but without us having to tell them how to complete the task.
Utility
One possibility of having agents complete tasks is to associate utilities with individual states and then have the agent aim to bring about states that maximise utility.
A task specification is a function: u : E -> R (where R is a real number). So we associate a value with every environment state.
But with this idea, what is the utility of a run (remember that a run is a set of states)… ?
Another possibility is to assign a utility to runs themselves:
u : R -> R.
This takes a more long term view than the previous approach.
Optimal Agents
The optimal agent in an environment is the one that maximises the expected utility.
Some agents cannot be implemented on some computers (some functions may require more memory than is available).
A bounded optimal agent is the agent, that can be implemented, that maximises expected utility.
Task Specification
0 or 1 can be assigned to previous runs, where 1 indicates that the agent succeeds, otherwise it fails.
These are known as predicate task specifications.
A task environment is a pair of enviornments and predicate task specifications.
Achievement and Maintenance Tasks
The two most common types of tasks are achievement and maintenance tasks.
An achievement task is one of the form: “achieve state of affairs”.
A maintenance task is one of the form: “maintain state of affairs”.
Related posts:
- Agent Architecture
- What is an Agent?
- Deductive Reasoning Agents
- An Introduction to Reverse Engineering
- User-Interface and Design Evaluation
Bookmarking
-
Stumble | Digg | del.icio.us | RSS
