Execution Architectures of RALPH-MEA
Influence diagrams provide one mechanism for combining probabilistic and
utility information. The state attributes become probabilistic nodes and
conditional probabilities become links. Action choices become decision
nodes and the utility function becomes a value node. The different knowledge types cover the same domain, though some
are more "compiled" than others. We can
define separate execution architectures, each using influence diagrams
formed through different combinations of knowledge types, and each
producing different behaviors.
Decision-Theoretic
Knowledge of types A,B, and C provide very fundamental information, taking
only single steps: first to elucidate the current state, then to find the
next state, and last to evaluate the next state. Using the Maximum Expected
Utility (MEU) principle to achieve rationality, this knowledge can be used to
pick locally optimal actions. However, calculation of MEU is
computationally intensive, often clashing with the limited response time afforded the system.
Goal-Based
Knowledge of types A and B again provide information about the next state,
but instead of using more general utility information, the goal-based EA
uses type F knowledge to as goal information. Thus there is no need to
invoke the MEU principle, since goal satisfaction is immediately
verifiable.
Action-Utility
Type E knowledge gives direct utility information of actions, after using
type A information to specify the current state. Then the MEU principle can
be invoked directly, without evaluation of a next state, increasing efficiency.
Condition-Action
Type D knowledge is even more efficient, immediately specifying the proper
action given a current state elucidated using type A statements. Thus no
evaluation of MEU is required, allowing high reactivity.
To return, press HOME.
To go to the next document, press NEXT.