Welcome to MLAP'08, a satellite Workshop of the 18th European Conference on Artificial Intelligence (ECAI 2008) on Machine Learning and Automated Planning.

Overview

Machine Learning (ML) is the area of Artificial Intelligence concerned with the design of computer programs that improve at a category of tasks with experience. It is a very broad field with many learning methodologies and numerous algorithms, which have been extensively exploited in the past to support planning systems in many ways.

First of all, since it is a usual case for seemingly different planning problems to present similarities in their structure, it is reasonable enough to believe that planning strategies that have been successfully applied to some problems in the past will be also effective for similar problems in the future. In this context, ML can assist planning systems in three main ways: a) to learn domain knowledge, b) to learn control knowledge and c) to learn optimization knowledge.

Domain knowledge is utilized by planners in pre-processing phases in order to either modify the description of the problem in a way that it will make it easier for solving or make the appropriate adjustments to the planner to best attack the problem. Control knowledge can be utilized during search in order to either solve the problem more efficiently or produce better plans. For example, the knowledge extracted from past examples can be used to refine the heuristic functions or create a guide for pruning non-promising branches. Most approaches in the past have focused on ways for learning control knowledge to improve efficiency since it is crucial for planners to have an informative guide during search. Finally, optimization knowledge is utilized after the production of an initial plan, in order to transform it in a new one that optimizes certain criteria, e.g. number of steps or resources usage.

On the other hand, planning is closely associated with Reinforcement Learning, a sub-area of Machine Learning where an agent learns from interaction with its environment. The main modelling concept in RL are Markov Decision Processes. In this context, planning can be used to take a model as input and produce or improve a policy for interacting with the modeled environment.

This workshop intends to provide a forum for researchers in the field of Machine Learning and Automated Planning, to discuss the above and other related topics regarding methodologies and systems that combine ideas from the two areas.

Topics

Possible topics of the workshop include (but are not limited to):

* Learning domain knowledge for planning
* Learning models to control the search
* Automatic optimization of planning systems
* Optimizing plans
* Reinforcement learning
* Fast replanning and plan adaptation
* Planning to sense
* Trading off planning time and plan quality,
* Multi-agent planning and learning
* Life-long learning of domain models
* Learning with very few training data,
* Theory of mixing inductive and deductive approaches to planning
* Learning for games (e.g., learning game evaluators at end games)
* Generalizing plans across similar domains
* Function approximation in planning
* Planning heuristics for exploration in reinforcement learning
* Applications of planning and learning
* Learning for planning in non-deterministic domains
* Learning for planning under uncertainty
* Mixed-initiative learning and planning systems