7 Relational Decision Theory

Relational decision theory, as presented in [7],
is implemented in Alchemy as an option when the user calls the `infer`
command. The input mln and evidence database are in the same format as when
running inference. The option `-decision` must be specified, and the output
is the optimal assignment to action atoms along with the utility achieved with
the optimal assignment.

In Markov Logic Decision Networks (MLDNs), utility weights can be specified alongside probability weights in the following manner:

`-2:20 Buys(x)`

Here, we want to have a high utility if someone buys something, but the prior
probability that someone actuall buys something is low.
Utilities may also be specified in isolation after `:` if you don't want a
weighted formula, but rather just want to specify a utility weight for an action
predicate:

`:-1 MarketTo(x)`

Here, we represent the cost of marketing something to someone.

Marc Sumner 2010-01-22