In this talk, I will focus on presenting our novel algorithms on anticipation, layered learning, and the integration of deliberative and reactive planning.
Firstly, I will present a state-action planning algorithm that allows for a team mate to "anticipate" the needs of other team mates. This anticipation strategic positioning algorithm computes the team member's moves that maximize the chances of successful collaboration, as a function of the dynamic state information.
Secondly, I will introduce layered learning as a technique to approach complex learning tasks. Learning applies at different levels of the task decomposition. The outcome of each level is efficiently used as the input for the next learning level. I will present three levels of learning.
Finally, I will introduce an algorithm for interleaving of deliberative and reactive planning suitable for real-time dynamic environments. Two main aspects are responsible for the success of the approach. First, the deliberative planner uses depth-bounded forward chaining guided by goal-based heuristics. Second, the real-time state space is discretized as a function of the average time that the deliberative planner needs to generate a plan.
I will illustrate the algorithms presented within the context of multi-agent robotic soccer. We participated at the RoboCup'98 competitions and came in first place in each of the three leagues we entered.