Comparision of the adaptation success of articulated artificial life forms in locomotion using various techniques

Locomotion is a favorable area to work on learning techniques. This situation arises from its capability to include all the elements of a difficult learning problem. Due to this fact, this study aims to apply and compare the effect of the usages of various learning techniques on an articulated artificial life form. This way the articulated structure tries to control its actions autonomously. Learning is implemented on an artificial environment which simulates nature?s physics laws in their basic form. Locomotion is a problem that requires adaptability. Thus instead of building a fixed locomotion control structure or providing control from outside of the structure, it is more desirable for the artificial life form to learn how to control its own actions. Therefore, this study makes use of three learning areas in order for the artificial life form to learn to act autonomously. The commonality of these three learning areas is that they have all been influenced by the way organisms operate. The learning areas worked on are evolutionary learning, supervised learning and reinforcement learning. Genetic Algorithms are used to realize evolutionary learning. Backpropagation Learning is applied on Artificial Neural Networks and Real Time Recurrent Learning is applied on Recurrent Neural Networks in order to realize supervised learning. Lastly, Q-Learning and Policy Gradient Reinforcement Learning are implemented to realize reinforcement learning. The conclusion of learning with Genetic Algorithms took much more time with respect to other learning techniques and, slow and distinctive movements were observed upon completion of the learning process. Real Time Recurrent Learning, which is one of the tried supervised learning techniques, produced better results than learning by Backpropagation. This technique eventuated fast. In addition, the artificial life form moved very quickly but directional deviations were observed frequently. Due to its inefficiency to deal with the defined large problem space, Q-learning turned out to be unsuccessful. The other reinforcement learning technique Policy Gradient Reinforcement Learning on the other hand, converged fast and resulted in large and distinctive movements. Only the pace of the locomotion turned out to be low. In conclusion, the implementations of all the learning areas resulted in some kind of locomotion that was visually above the threshold. The results of the learning techniques mostly differed on speed of learning, pace of locomotion and change in direction. The directional diversions of Real Time Recurrent Learning are aimed to be reduced in future studies. Additionally, the combination of Genetic Algorithms and Policy Gradient Reinforcement Learning is offered as another future work. Keywords: Artificial Life, Locomotion, Evolutionary Learning, Supervised Learning, Reinforcement Learning