Abstract

Deep reinforcement learning for interactive multimodal robots is attractive for endowing machines with trainable skill acquisition. But this form of learning still represents several challenges. The challenge that we focus in this paper is effective policy learning. To address that, in this paper we compare the Deep Q-Networks (DQN) method against a variant that aims for stronger decisions than the original method by avoiding decisions with the lowest negative rewards. We evaluated our baseline and proposed algorithms in agents playing the game of Noughts and Crosses with two grid sizes (3x3 and 5x5). Experimental results show evidence that our proposed method can lead to more effective policies than the baseline DQN method, which can be used for training interactive social robots.