# Run 2024-9-23 (Discarded)

* 2024-9-23
* Start the training.
* The version is v0.8.0.
* learning rate = 0.005
* batch size = 256
* The network size is 6bx96c.
* area scoring and territory scoring.

</br>

* 2024-9-25
* Played 275k games.
* Accumulate around $0.473\times 10^{8}$ 20bx256c eval queries.
* The strengh is better than Leela Zero with [91](https://zero.sjeng.org/networks/b3b00c6d75b4e74946a97b88949307c9eae2355a88f518ebf770c7758f90e357.gz) weights. Elo different is +42.
* Sayuri: ```-w current_weights -t 1 -b 1 -p 800 --cpuct-init 0.5 --use-optimistic-policy --random-moves-factor 0.1 --random-moves-temp 0.8```
* Leela Zero: ```-w 091.gz --noponder -v 800 -g -t 1 -b 1 --timemanage off --randomcnt 30 --randomtemp 0.8```
* Game result (played 569 games with Leela Zero):

        Name             black won       white won    total (win-rate)
        
        Sayrui v0.8.1      150              169         319  (56.06%)
        Leela Zero 0.17    115              135         250  (43.94%)

</br>

* 2024-9-26
* Played 340k games.
* Accumulate around $0.732\times 10^{8}$ 20bx256c eval queries.
* The strengh is better than Leela Zero with [92](https://zero.sjeng.org/networks/ae205d8b957e560c19cc2bc935a8ea76d08dd9f88110ea783d50829bdca45329.gz) weights. Elo different is +40.
* Sayuri: ```-w current_weights -t 1 -b 1 -p 800 --cpuct-init 0.5 --use-optimistic-policy --random-moves-factor 0.1 --random-moves-temp 0.8```
* Leela Zero: ```-w 092.gz --noponder -v 800 -g -t 1 -b 1 --timemanage off --randomcnt 30 --randomtemp 0.8```
* Game result (played 734 games with Leela Zero):

        Name             black won       white won    total (win-rate)
        
        Sayrui v0.8.1      202              207         409  (55.72%)
        Leela Zero 0.17    164              161         325  (44.28%)

* The strengh is weak than Leela Zero with [95](https://zero.sjeng.org/networks/5b90bd32ccc835d8e08d41970d39753e5732413d75e8f4035bebb5f1da69fb87.gz) weights. Elo different is -24.
* Sayuri: ```-w current_weights -t 1 -b 1 -p 800 --cpuct-init 0.5 --use-optimistic-policy --random-moves-factor 0.1 --random-moves-temp 0.8```
* Leela Zero: ```-w 095.gz --noponder -v 800 -g -t 1 -b 1 --timemanage off --randomcnt 30 --randomtemp 0.8```
* Game result (played 611 games with Leela Zero):

        Name             black won       white won    total (win-rate)
        
        Sayrui v0.8.1      129              155         284  (46.48%)
        Leela Zero 0.17    150              177         327  (53.52%)

</br>

* 2024-9-30
* Played 590k games.
* Accumulate around $1.712\times 10^{8}$ 20bx256c eval queries.
* **Halt the 6bx96c training.**
* The strengh is as same as Leela Zero with [110](https://zero.sjeng.org/networks/4c7dd084662bfe356e572d2dca033a8e0dd30d1542bb252083ffc899cdbcb52e.gz) weights. Elo different is +9.
* Sayuri: ```-w current_weights -t 1 -b 1 -p 800 --cpuct-init 0.5 --use-optimistic-policy --random-moves-factor 0.1 --random-moves-temp 0.8```
* Leela Zero: ```-w 110.gz --noponder -v 800 -g -t 1 -b 1 --timemanage off --randomcnt 30 --randomtemp 0.8```
* Game result (played 1000 games with Leela Zero):

        Name             black won       white won    total (win-rate)
        
        Sayrui v0.8.1      239              274         513  (51.30%)
        Leela Zero 0.17    226              261         487  (48.70%)

* The strengh is weak than Leela Zero with [116](https://zero.sjeng.org/networks/39d465076ed1bdeaf4f85b35c2b569f604daa60076cbee9bbaab359f92a7c1c4.gz) weights. Elo different is -14.
* Sayuri: ```-w current_weights -t 1 -b 1 -p 800 --cpuct-init 0.5 --use-optimistic-policy --random-moves-factor 0.1 --random-moves-temp 0.8```
* Leela Zero: ```-w 116.gz --noponder -v 800 -g -t 1 -b 1 --timemanage off --randomcnt 30 --randomtemp 0.8```
* Game result (played 1000 games with Leela Zero):

        Name             black won       white won    total (win-rate)
        
        Sayrui v0.8.1      227              253         480  (48.00%)
        Leela Zero 0.17    247              273         520  (52.00%)

</br>

* 2024-10-1
* Start the 10bx128c training.
* Current replay buffer is 112500 games.

</br>

* 2024-10-4
* Played 720k games (10bx128c played 130k games).
* Accumulate around $3.178\times 10^{8}$ 20bx256c eval queries.
* The strengh is better than Leela Zero with [117](https://zero.sjeng.org/networks/ba748d402af1bdd101212cbd217025dee866b3fcc996bd16d1a3134d5591a501.gz) weights. Elo different is +20.
* Sayuri: ```-w current_weights -t 1 -b 1 -p 800 --cpuct-init 0.5 --use-optimistic-policy --random-moves-factor 0.1 --random-moves-temp 0.8```
* Leela Zero: ```-w 117.gz --noponder -v 800 -g -t 1 -b 1 --timemanage off --randomcnt 30 --randomtemp 0.8```
* Game result (played 707 games with Leela Zero):

        Name             black won       white won    total (win-rate)
        
        Sayrui v0.8.1      190              184         374  (52.90%)
        Leela Zero 0.17    169              164         333  (47.10%)

</br>

* 2024-10-7
* Played 925k games (10bx128c played 335k games).
* Accumulate around $5.466\times 10^{8}$ 20bx256c eval queries.
* Drop the learning rate to 0.0025 (from 0.005).
* Current replay buffer is 131250 games.
* Strength should be slightly stronger than LZ125.

</br>

* 2024-10-7
* Played 935k games (10bx128c played 345k games).
* Accumulate around $5.578\times 10^{8}$ 20bx256c eval queries.
* The strengh is as same as Leela Zero with [130](https://zero.sjeng.org/networks/18e6a6c53d61359a922b36a5f3fb078aa7514cfe0050ccf3cae37f8584154d89.gz) weights. Elo different is +9.
* Sayuri: ```-w current_weights -t 1 -b 1 -p 800 --cpuct-init 0.5 --use-optimistic-policy --random-moves-factor 0.1 --random-moves-temp 0.8```
* Leela Zero: ```-w 130.gz --noponder -v 800 -g -t 1 -b 1 --timemanage off --randomcnt 30 --randomtemp 0.8```
* Game result (played 1000 games with Leela Zero):

        Name             black won       white won    total (win-rate)
        
        Sayrui v0.8.1      242              271         513  (52.90%)
        Leela Zero 0.17    229              258         487  (47.10%)

</br>

* 2024-10-11
* Played 1140k games (10bx128c played 550k games).
* Accumulate around $7.886\times 10^{8}$ 20bx256c eval queries.
* **Halt the 10bx128c training.
* The strengh is better than Leela Zero with [135](https://zero.sjeng.org/networks/05b723d0c9e1c10abf18e9ba3e55c378de42e364cfe8a506d425f2af8ca204e8.gz) weights. Elo different is +15.
* Sayuri: ```-w current_weights -t 1 -b 1 -p 800 --cpuct-init 0.5 --use-optimistic-policy --random-moves-factor 0.1 --random-moves-temp 0.8```
* Leela Zero: ```-w 135.gz --noponder -v 800 -g -t 1 -b 1 --timemanage off --randomcnt 30 --randomtemp 0.8```
* Game result (played 1000 games with Leela Zero):

        Name             black won       white won    total (win-rate)
        
        Sayrui v0.8.1      246              276         522  (52.90%)
        Leela Zero 0.17    224              254         487  (47.10%)

* The strengh is weak than Leela Zero with [140](https://zero.sjeng.org/networks/521b086873ab560bbd34bab3b5d98a30efd5a4292e4c589aca3d9a3d03165ecf.gz) weights. Elo different is -35.
* Sayuri: ```-w current_weights -t 1 -b 1 -p 800 --cpuct-init 0.5 --use-optimistic-policy --random-moves-factor 0.1 --random-moves-temp 0.8```
* Leela Zero: ```-w 140.gz --noponder -v 800 -g -t 1 -b 1 --timemanage off --randomcnt 30 --randomtemp 0.8```
* Game result (played 546 games with Leela Zero):

        Name             black won       white won    total (win-rate)
        
        Sayrui v0.8.1      106              140         246  (45.05%)
        Leela Zero 0.17    133              167         300  (54.95%)

</br>

* 2024-10-15
* Start the 15bx192c training.
* Current learning rate is 0.0025.
* Current replay buffer is 168750 games.

</br>

* 2024-10-20
* Played 1320k games (15bx192c played 180k games).
* Accumulate around $1.4773\times 10^{9}$ 20bx256c eval queries.
* The strengh is better than Leela Zero with [155](https://zero.sjeng.org/networks/834f35fa55ef2f46d40021b2cffd8f30a7f389e7664ff38aaf039e4c2e17265c.gz) weights. Elo different is +31.
* Sayuri: ```-w current_weights -t 1 -b 1 -p 800 --cpuct-init 0.5 --use-optimistic-policy --random-moves-factor 0.1 --random-moves-temp 0.8```
* Leela Zero: ```-w 155.gz --noponder -v 800 -g -t 1 -b 1 --timemanage off --randomcnt 30 --randomtemp 0.8```
* Game result (played 570 games with Leela Zero):

        Name             black won       white won    total (win-rate)
        
        Sayrui v0.8.1      146              164         310  (54.39%)
        Leela Zero 0.17    121              139         260  (45.61%)

* The strengh is as same as Leela Zero with [157](https://zero.sjeng.org/networks/d351f06e446ba10697bfd2977b4be52c3de148032865eaaf9efc9796aea95a0c.gz) weights. Elo different is +3.
* Sayuri: ```-w current_weights -t 1 -b 1 -p 800 --cpuct-init 0.5 --use-optimistic-policy --random-moves-factor 0.1 --random-moves-temp 0.8```
* Leela Zero: ```-w 157.gz --noponder -v 800 -g -t 1 -b 1 --timemanage off --randomcnt 30 --randomtemp 0.8```
* Game result (played 615 games with Leela Zero):

        Name             black won       white won    total (win-rate)
        
        Sayrui v0.8.1      135              175         310  (50.41%)
        Leela Zero 0.17    132              173         305  (49.59%)


</br>

* 2024-10-24
* Played 1445k games (15bx192c played 305k games).
* Full Visits = 400 (from 150)
* Fast Visits = 125 (from 50)
* Gumbel Visits = 55 (from 32)
* Resign Visits = 75 (from 55)

</br>

* 2024-11-15
* Played 1965k games (15bx192c played 825k games).
* Accumulate around $5.569\times 10^{9}$ 20bx256c eval queries.
* The strength is still weaker than ELFv0. Best weights is only 46% win-rate. The progression is bad so we decide to halt 15bx192c training.
* We drop the learning rate at
* 15bx192c played 365k games | at 2024-10-27 | drop lr to ```0.002```
* 15bx192c played 475k games | at 2024-10-31 | drop lr to ```0.00125```
* 15bx192c played 730k games | at 2024-11-11 | drop lr to ```0.001```