Distributed control of multi-actor systems: Reinforcement signal by Shannon’s entropy

Youcef Zennir , Denis Pomorski (1)

1 LAGIS - Laboratoire d'Automatique, Génie Informatique et Signal

This paper presents a control of multi-actor systems (or multi-sensors) functioning in an unknown environment. These actors are reactive entities able to react to the stimuli coming from the environment and to choose between several actions. In order to improve their behaviour (i.e. in order to choose the good action) in the course of time, the system multi-actors must be able to use reinforcement learning. This signal of reinforcement is, until now, a signal whose values are a priori defined. We propose to raise this "a priori" while using the Shannon’s entropy to measure the coherence of the choice of the action by the transformation of the reinforcement signal table. This stage, of local training will allow the improvement of the control of the global system and coordination between the various actors. The results of the simulation show that the actor can learn to control its trajectory efficiently.

Mots clés: Reinforcement learning; distributed control; Q-learning; multi-actor systems

Domaines: Automatique; Apprentissage [cs.LG]; Théorie de l'information [cs.IT]

Youcef Zennir, Denis Pomorski. Distributed control of multi-actor systems: Reinforcement signal by Shannon’s entropy. The 24th European Modelling and Simulation Symposium (Simulation in Industry), EMSS’2012, Sep 2012, Vienna, Austria. ⟨hal-01509724⟩