Post by steadyeddie on Dec 18, 2016 6:28:37 GMT -8
Adding RAVE to SteadyEddie definitely made it better, but I find it all a bit of a dark art.
I think you are supposed to alter scores of nodes associated with moves. Of course, that's not much good if you are playing chess, or dots and boxes, as it's the states that are important, but it definitely works to some degree anyway.
But even if you are doing just move-based, there are questions about how you use the information.
Should you select new, non-visited, nodes based on RAVE scores? Experimentation suggests, yes, you should. But should you use RAVE if some children have had no visits, and go for more attractive nodes, based on RAVE? The answer appears to be no. Which is not what I expected.
Then you should update the scores in the tree and you get new information. But how should that be weighted against actual rollout data? My experimentation has shown not as a percentage (like 10% rave, 90% rollout), so I use weight like it counts as 10 rollouts. I also experimented with it counting towards visits, but that didn't seem to work out.
I've also experimented with using states to control the RAVE scores. I track state changes in the rollouts, and keep stats, and use the differences between the parent and the child to give it an effective rave score. You'd have thought that would have a positive effect, but not so much :-(
I think you are supposed to alter scores of nodes associated with moves. Of course, that's not much good if you are playing chess, or dots and boxes, as it's the states that are important, but it definitely works to some degree anyway.
But even if you are doing just move-based, there are questions about how you use the information.
Should you select new, non-visited, nodes based on RAVE scores? Experimentation suggests, yes, you should. But should you use RAVE if some children have had no visits, and go for more attractive nodes, based on RAVE? The answer appears to be no. Which is not what I expected.
Then you should update the scores in the tree and you get new information. But how should that be weighted against actual rollout data? My experimentation has shown not as a percentage (like 10% rave, 90% rollout), so I use weight like it counts as 10 rollouts. I also experimented with it counting towards visits, but that didn't seem to work out.
I've also experimented with using states to control the RAVE scores. I track state changes in the rollouts, and keep stats, and use the differences between the parent and the child to give it an effective rave score. You'd have thought that would have a positive effect, but not so much :-(