Adaptive intelligence is the ability of a system to respond quickly and effectively to changes in its environment during its lifetime. Responding quickly to change can be greatly beneficial in many domains, where unpredictable changes require generalizing adaptive behavior. Although there are many online reinforcement learning systems, manually selecting the proper representation can be challenging, and often leads to suboptimal solutions. By using evolution to automate the search for an appropriate representation, we are able to evolve agents better able to learn. We will examine the applicability of this concept by using real time NeuroEvolving Augmenting Topologies (rtNEAT) - an evolutionary algorithm for exploring the topological and synaptic weight space of neural networks - to evolve a function approximator for the Q-learning temporal difference method using an adaptive learning rate to more effectively adjust to dynamic conditions. This algorithm, real time NeuroEvolving Augmenting Topologies + Q-learning with Adaptive Learning Rate (rtNEAT+QwALR) will be compared in the dangerous foraging domain with adaptive assignments against rtNEAT, and rtNEAT+Q.

PDF

Technical Report HR-09-04, Department of Computer Science, The University of Texas at Austin., 2009.

@techreport{tucker:ugthesis09, title={Evolving Adaptive Intelligence: Using NeuroEvolution with Temporal Difference Methods in the Game Domain}, author={Nathaniel Tucker}, number={HR-09-04}, institution={Department of Computer Science, The University of Texas at Austin.}, type={Undergraduate Honors Thesis}, url="http://www.cs.utexas.edu/users/ai-lab/?tucker:ugthesis09", year={2009} }