Rule value reinforcement learning for cognitive agents

ABSTRACT

RVRL (Rule Value Reinforcement Learning) is a new algorithm which extends an existing learning framework that models the environment of a situated agent using a probabilistic rule representation. The algorithm attaches values to learned rules by adapting reinforcement learning. Structure captured by the rules is used to form a policy. The resulting rule values represent the utility of taking an action if the rule's conditions are present in the agent's current percept. Advantages of the new framework are demonstrated, through examples in a predator-prey environment.

REFERENCES

• 1: Child, C. and Stathis, K. "Learning to Act with RVRL agents". City University Technical Report. (2006).
• 2: Child, C. and Stathis, K. "The Apriori Stochastic Dependency Detection (ASDD) Algorithm for Learning Stochastic Logic Rules", In Proceedings of the 4th International Workshop on Computation Logic in Multi-agent Systems (CLIMA-04), J. Dix, J. Leiter (Eds), Florida, Jan. (2004).
• 3: Pasula, H. M. Zettlemoyer, L. S. and Kaelbling, L. P. "Learning Probabilistic Relational Planning Rules." Proceedings of the Fourteenth International Conference on Automated Planning and Scheduling, ICAPS, 73--82, (2004).
• 4: Richard S. Sutton , Andrew G. Barto, Introduction to Reinforcement Learning, MIT Press, Cambridge, MA, 1998
• 5: Watkins, C. J. C. H. "Learning from Delayed Rewards." PhD thesis, Cambridge University, (1989).

INDEX TERMS

Primary Classification:
  I. Computing Methodologies
  I.2 ARTIFICIAL INTELLIGENCE
      I.2.1 Applications and Expert Systems

General Terms:
Algorithms

Keywords:
action, environment, perception, planning, probabilistic logic, reinforcement learning, situated agents, stochastic