Multi-Stage Temporal Difference Learning for 2048 - Technologies and Applications of Artificial Intelligence - page 370

3 Our TD Learnin ng Method for 2048

This paper designs our TD
our method, we first chang
MS-TD learning in Subsect
in Subsection 3.4 and some

learning method for 2048 based on the method in [17]
ge the n-tuple network in Subsection 3.1, and then prop
tion 3.2. The experiments for MS-TD learning are descri
issues are discussed in Subsection 3.5.

]. In
pose
ibed

3.1 New N-Tuple Netw work

In this paper, we use the tu
mirrored tuples. In brief, w
shape as shown in Fig. 4 (
tuples. The number of featu

uples shown in Fig. 4 (b) as well as all of their rotated
we change 4-tuples (1x4 lines) to 6-tuples in a green kn
(a). Apparently, the new 6-tuples cover all the origina
ures increases only by a factor of two or so.

and
nife
al 4-

cores in TD learning with different n-tuple networks

Fig. 5. Average s

scores in TD learning with different n-tuple networks

Fig. 6. Maximum