- Abstract:
-
We introduce a model of animat reinforcement learning where an associative learning element is placed between the animat's reinforcement learning component and its internal reinforcement functions. This element forms an impression of the sensory stimuli present near a goal and uses it to make an initial estimate of the value of newly discovered state-action pairs in tasks where reward is necessarily delayed.
We then describe the implementation of Peaches 'n Cream, a simulated robot that implements one version of the model where behaviour-based reinforcement learning is used in a puck foraging task. The results suggest that once given an initial simple task to learn from, the resulting associations significantly speed up learning in a later, more complex task, and provide further evidence that other learning methods may be used in conjunction with reinforcement learning to make it feasible for situated agents.
- Links To Paper
- No links available
- Bibtex format
- @InProceedings{EDI-INF-RR-0729,
- author = {
George Konidaris
and Gillian Hayes
},
- title = {Estimating future reward in reinforcement learning animats using associative learning},
- book title = {From Animals to Animats - Proceedings of the eighth international conference on the simulation of adaptive behavior (SAB8)},
- publisher = {MIT Press},
- year = 2004,
- month = {Jul},
- volume = {8},
- pages = {297-304},
- }
|