As the semantics, language bias and even the logic programming paradigm of ILP frameworks is widely varied, and we would like this competition to be open to any system, the format of our learning task will not be aimed at any particular learning setting.
The competition will have two main tracks: probabilistic and non-probabilistic. You may enter one or both of these tracks. While simple tasks will consist of observational predicate learning of definite logic programs, there are also several advanced sub-categories. Do not worry if your system is unable to cope with every type of task, as we expect different systems to have different strengths! The advanced categories are: non-observational predicate learning, predicate invention, non-monotonic and recursion.
The problem domain will be of an agent learning about its environment. In each problem instance, the learners will be given a map (as a set of facts and definite rules describing attributes of the cells) and a set of traces through the map taken by the agent along with complete sets of valid moves for each trace. An example of a trace through a map is shown below. The agent is labeled A and the valid moves are labeled VM. You can use the time selector to move the agent through the map.
The agent may move to any adjacent cell which is not already full so long as there is no wall between the cell and the agent's current position.
In addition to the set of example traces with valid moves, your system will also be given a set of test traces with no valid moves. You should use the example traces to find a hypothesis defining "valid move". It is then your system's task to decide whether or not the test traces are valid (in the probabilistic case, you should return the probability of each trace being valid).
Task Format and Entry
You can find more information about how to enter the competition and how to enter on the instructions page.