Going Further: Writing AI Algorithms
The common library makes heavy use of three interfaces that cover a wide range of algorithms.
The following questions cover these three interfaces:
Can the algorithm be implemented as a pure function?
Can the algorithm be called multiple times to return alternative solutions to the same problem?
Does the algorithm have parameters?
Function Interface
The core library provides a simple interface for AI algorithm which takes a term as input and returns a term as a result.
To implement a learning algorithm, your input may be a (labeled) data set and your output would a model (such as a decision tree).
For a reasoning algorithm, the input could contain a knowledge base and a query and the output could be one or more answers to the query.
In case of planning, the input would contain an initial state, a goal and a set of operators and the output is a plan or a symbol indicating failure in case no plan exists.
Initializable
Some algorithms behave like iterators in the sense that they can be asked for further solutions. In case of planning, the first plan may be rejected by a user. In such cases we may want to continue search (possibly after adding some additional constraints).
For this purpose, the intilizable interface exists to allow functions to be set to a clear initial state and then support multiple calls of the actual function (e.g., to get the next solution).
Example: search algorithms are initialized with a set of possible start nodes that should be used as origins of the search.
Configuration
Many algorithms include parameters that may influence performance on specific problems. For this purpose the configurable interface exists which allows a common way to configure any AIDDL common algorithm through a single AIDDL term (which can contain an arbitrary amount of parameters if needed).
This covers learning rates, neural network layouts, thresholds, or maximum iterations for stochastic search algorithms.
Other Interfaces / Traits
Graph Search
Tree Search
Generic trait that allows searching in a tree
Learner
Extension of the function interface which adds methods for fitting and prediction.