Integrated modeling and control based on reinforcement learning and dynamic programming
NIPS-3 Proceedings of the 1990 conference on Advances in neural information processing systems 3
C4.5: programs for machine learning
C4.5: programs for machine learning
Introduction to Reinforcement Learning
Introduction to Reinforcement Learning
Scaling Reinforcement Learning toward RoboCup Soccer
ICML '01 Proceedings of the Eighteenth International Conference on Machine Learning
Relational Reinforcement Learning
ICML '98 Proceedings of the Fifteenth International Conference on Machine Learning
ICML '04 Proceedings of the twenty-first international conference on Machine learning
AAAI'05 Proceedings of the 20th national conference on Artificial intelligence - Volume 2
Learning planning rules in noisy stochastic worlds
AAAI'05 Proceedings of the 20th national conference on Artificial intelligence - Volume 2
Top-down induction of first-order logical decision trees
Artificial Intelligence
Probabilistic relational planning with first order decision diagrams
Journal of Artificial Intelligence Research
| Hi-index | 0.00 |
Many reinforcement learning domains are highly relational. While traditional temporal-difference methods can be applied to these domains, they are limited in their capacity to exploit the relational nature of the domain. Our algorithm, AMBIL, constructs relational world models in the form of relational Markov decision processes (MDPs). AMBIL works backwards from collections of high-reward states, utilizing inductive logic programming to learn their preimage, logical definitions of the region of state space that leads to the highreward states via some action. These learned preimages are chained together to form an MDP that abstractly represents the domain. AMBIL estimates the reward and transition probabilities of this MDP from past experience. Since our MDPs are small, AMBIL uses value-iteration to quickly estimate the Q-values of each action in the induced states and determine a policy. AMBIL is able to employ complex background knowledge and supports relational representations. Empirical evaluation on both synthetic domains and a sub-task of the RoboCup soccer domain shows significant performance gains compared to standard Q-learning.