Database Reference
In-Depth Information
AgentSettings
Settings
AgentSpecification
Apply Data
Learn Data
Agent
Action(s)
Learn
Environment
Information
Environment
Apply
Fig. 12.15 Main interfaces of agent
covered through new data which is handed to the agent in a so-called learn step.
Furthermore, the agent needs to be triggered to take an action. This is in contrast
to some common agent theories, where an agent decides by himself or herself when
to act.
12.2.1.1 Agent
The class
Agent
which represents the root of all agent implementations is abstract.
The dataflow is shown in Fig.
12.15
. If we compare it with the dataflow of data
mining of Fig.
12.11
, we see that the main idea is to join the classes of the
algorithm and the model into one class - the agent. This is because unlike as in
classical data mining an agent does both learning and application, often combined.
As we see, the agent parameters are passed through the agent settings on agent-
type level and agent specification on algorithm level - very similar to mining
settings and mining algorithm specification of the mining process. Additional
information like product-specific
master
data required in the specific agent appli-
cations can be specified through the environment information which, however, is
mandatory.
EnvironmentInformation
is basically a hash map of mining input streams where
the keys are the names of the streams. Thus, each stream represents a table. For
example, there may be a key
product
for a mining input stream of all product
informations of a web shop and
regions
for tabular informations about different
regions. The inclusion of
EnvironmentInformation
into the
Agent
package is impor-
tant because unlike data mining, which almost always works on a flat table, agents
are often more complex and work on nested schemas. For example, XELOPES
contains business-oriented packages of disposition and price optimization as well as
a reinforcement learning package described in the next section.
