X-Git-Url: https://git.auder.net/?p=rpsls-bot.git;a=blobdiff_plain;f=README.md;h=2bf8a5fe802fff30cc3010993016f71a3a8be4ef;hp=c72ae2de7d5f0985aad909f487b8b488d601b5fc;hb=HEAD;hpb=6fac12b0c92df43cf7c8d26c1c5b480b3f8940fb

diff --git a/README.md b/README.md
index c72ae2d..2bf8a5f 100644
--- a/README.md
+++ b/README.md
@@ -6,8 +6,26 @@ The rules are given by Sheldon in episode 8 of season 2 of TBBT (The Big Bang Th
 
 ---
 
-[Online demo](https://auder.net/rpsls/")
+[Online demo](https://auder.net/rpsls/)
 
 Winning should be difficult after a few dozens of rounds, because it's hard to play at random.
 
-Setting "draw is lost" and/or increasing number of inputs can improve bot level.
+Setting "winner bot" and/or increasing memory can improve bot level.
+
+---
+
+## Technical details
+
+Each potential choice is linked to all outputs in a (neural) network, for
+each input in memory. We thus have size of memory x (number of choice)^2 links.
+To select a move, the bot computes the sum of all links weights from an activated choice
+(that is to say, the value of a memory cell) to each output.
+The output with biggest weights sum wins: the move is played.
+
+The reward is then determined from human move: -1 for a loss, 0 for a draw
+(except if "winner bot" is selected, in which case a draw = a loss) and 1 for a win.
+Weights on the active links are updated positively or negatively depending on reward sign.
+All weights are initialized to zero, and since some time is required for learning
+the first moves in the game would be quite random.
+
+See RPS\_network\_2.svg file for an illustration with memory=2 and simple RPS.