Chasing the Sun: In quest of Human Intelligence

It is by studying human intelligence, that we achieved the creation of machines able to perform intelligent behaviors, i.e. learn from experience to solve problems. These studies have included how humans have learnt to distinguish complex features and to infer the status and place of objects within an environment, or to interact with other humans for solving complex problems like management of large organizations or even cooperative activities like sport competitions. To study these aspects, this work addresses the topics of unbalanced object classification and machine cognitive cooperation. The research is focused on investigating if virtual agents are capable of cooperating between them by inferring each others actions without explicit communication; and how to improve the performance of object classification in presence of critical datasets. The novelty of this work is two-fold: the implementation of transfer learning, computer vision filtering techniques and key performance metric engineering to deal with highly unbalanced and artifact-polluted datasets; and the design of a novel cooperation paradigm where distinct agents learn to successfully modify their policies to reach a common goal. We experimented both algorithms in two different test environments. First, we tested the new classification algorithm in an high challenging industrial context where defect detection should be extremely accurate, the new algorithms achieve an accuracy of 96.30\%. Secondly, the cooperation algorithm was checked to prove the ability of the agent in solving a complex maze while adapting their policies in the control of the same avatar both in discrete and continuous environments and mimicking what the human counterpart usually does in similar settings. The system, methodology and conclusions found can also be easily extended to other domains. The approach to defect classification can be applied to all production phases where visual inspection is used to assess the presence of specific characteristics within the analyzed element. The Machine Cognitive Cooperation Reinforcement Learning task can be applicable to several real-life scenarios where observation and reward decomposition and the modularization of the cognitive processing can result advantageous for solving complex environments