Optimistic Temporal Difference Learning for <i>2048</i>

Temporal difference (TD) learning and its variants, such as multistage TD (MS-TD) temporal coherence (TC) learning, have been successfully applied to 2048. These methods rely on the stochasticity of environment 2048 for exploration. In this paper, we propose employ optimistic initialization (OI) encourage exploration 2048, empirically show that quality is significantly improved. This approach optimistically initializes feature weights very large values. Since tend be reduced once states are visited, agents explore those which unvisited or visited few times. Our experiments both TC with OI improve performance. As a result, network size required achieve same performance reduced. With additional tunings expectimax search, tile-downgrading technique, our design achieves state-of-the-art performance, namely an average score 625 377 rate 72% reaching 32768 tiles. addition, sufficiently tests, 65536 tiles reached at 0.02%.