Behaviour design in microrobots: hierarchical reinforcement learning under resource constraints

In order to verify models of collective behaviors of animals, robots could be manipulated to implement the model and interact with real animals in a mixed-society. This thesis describes design of the behavioral hierarchy of a miniature robot, that is able to interact with cockroaches, and participates in their collective decision makings. The robots are controlled via a hierarchical behavior-based controller in which, more complex behaviors are built by combining simpler behaviors through fusion and arbitration mechanisms. The experiments in the mixed-society confirms the similarity between the collective patterns of the mixed-society and those of the real society. Moreover, the robots are able to induce new collective patterns by modulation of some behavioral parameters. Difficulties in the manual extraction of the behavioral hierarchy and inability to revise it, direct us to benefit from machine learning techniques, in order to devise the composition hierarchy and coordination in an automated way. We derive a Compact Q-Learning method for micro-robots with processing and memory constraints, and try to learn behavior coordination through it. The behavior composition part is still done manually. However, the problem of the curse of dimensionality makes incorporation of this kind of flat-learning techniques unsuitable. Even though optimizing them could temporarily speed up the learning process and widen their range of applications, their scalability to real world applications remains under question. In the next steps, we apply hierarchical learning techniques to automate both behavior coordination and composition parts. In some situations, many features of the state space might be irrelevant to what the robot currently learns. Abstracting these features and discovering the hierarchy among them can help the robot learn the behavioral hierarchy faster. We formalize the automatic state abstraction problem with different heuristics, and derive three new splitting criteria that adapt decision tree learning techniques to state abstraction. Proof of performance is supported by strong evidences from simulation results in deterministic and non-deterministic environments. Simulation results show encouraging enhancements in the required number of learning trials, robot's performance, size of the learned abstraction trees, and computation time of the algorithms. In the other hand, learning in a group provides free sources of knowledge that, if communicated, can broaden the scales of learning, both temporally and spatially. We present two approaches to combine output or structure of abstraction trees. The trees are stored in different RL robots in a multi-robot system, or in the trees learned by the same robot but using different methods. Simulation results in a non-deterministic football learning task provide strong evidences for enhancement in convergence rate and policy performance, specially in heterogeneous cooperations.

Related material