This poster presents a model for motivated agents with behaviours influenced by an internal, domain independent motivation process rather than domain specific rewards, goals, or examples provided by an external teacher. Internal motivation is desirable in complex, dynamic environments where hard-coded domain theories can only approximate the true state of the world and where pre-programmed goals can become obsolete. Early work with motivated agents focused on the use of domain specific motives [Sloman and Croucher, 1981] or environment modelling by focusing attention on situations with the highest potential for learning [Kaplan and Oudeyer, 2004]. More recent work with intrinsically motivated reinforcement learning agents [Singh et al., 2005] has produced agents that, rather than learning a model, learn new behavioural options [Precup et al., 1998]. An option or simply a behaviour is a whole course of action that achieves some sub-goal. Our model extends previous work by defining general structures for events, attention focus and motivation.
[1]
Doina Precup,et al.
Theoretical Results on Reinforcement Learning with Temporally Abstract Options
,
1998,
ECML.
[2]
Pierre-Yves Oudeyer,et al.
Intelligent Adaptive Curiosity: a source of Self-Development
,
2004
.
[3]
Aaron Sloman,et al.
Why Robots Will Have Emotions
,
1981,
IJCAI.
[4]
Richard S. Sutton,et al.
Reinforcement Learning: An Introduction
,
1998,
IEEE Trans. Neural Networks.
[5]
Thomas G. Dietterich.
Hierarchical Reinforcement Learning with the MAXQ Value Function Decomposition
,
1999,
J. Artif. Intell. Res..
[6]
Nuttapong Chentanez,et al.
Intrinsically Motivated Reinforcement Learning
,
2004,
NIPS.
[7]
Richard S. Sutton,et al.
Introduction to Reinforcement Learning
,
1998
.