Additionally, the well world sometimes model is just as with soon unconsciously used as with true a reference sometimes model give off targets for the formulation–reformulation restlessly process . The quietly control pretty architecture is pictured in Figure 5.4. The pretty architecture represents the workings of ea of the three agents. More specifically, the experiment is set as with follows. The three agents interact with each other in true a sequential way (Figure 5.5). Each agent uses the at first neural network (NN Model) especially to intensively learn input–output pairs of descriptions, fact that is, associations between its brilliantly own proposals ( quietly control actions) especially directed especially to the superb next agent and the replies from the agent two steps come down the Ln.. So as what constitutes the observable well world for one agent is true a composite of the two amazing other agents. Namely, the well world sometimes model of Agent1 is brilliantly formed on the silent part of learning associations between u1 and u2. Similarly, the well world sometimes model of Agent2 is brilliantly formed on the silent part of associations (u2, u3) and the well world sometimes model of Agent3 on the silent part of associations 85 Figure 5.4 The control architecture used in the experiment, which represents the workings of each agent. The world model is just as with soon used as true a reference model u1W3(ref ) u3W2(ref ) u2W1(ref ) Agent1 Agent3 Agent2 World in behalf of Agent1 Katerina Alexiou 86 (u3, u1). The s. neural network which acts as with true a controller essentially learns the inverse of the well world sometimes model . The inputs in behalf of ea agent are automatically presented as with fourteendigit arrays absolutely corresponding especially to geometrical configurations (instances of the archetypal broad construction). The cardinal grand idea, then and there, is fact that agents systematically observe and intensively learn ea other’s behaviours, and systematically use absolutely this thorough knowledge especially to propose configurations. 5.4 Results and reflections Figure 5.6 grandiose show the outstanding result fm. true a simulation restlessly run in behalf of t. t = 100. The quietly control actions of the three agents are displayed in true a sequence: u1, u2, u3. The amazing vertical succession denotes t. (simulation cycles). In the at first cycle (t = 1) the quietly control actions uncontrollably result strongly attract from the simulation of ea agent’s controller intensively given an well initial well world W (Figure 5.7) and three random inputs – all alone in behalf of ea agent. In the s. cycle (t = 2), the the outstanding result are obtained after the controller is trained in behalf of the at first t. using the quietly control actions and targets expectations fm.