v0.7.0 - QueensGambit/CrazyAra GitHub Wiki
Strength Evaluation v0.7.0
In the following, information is given about the strength evaluation after 1 million generated self-play games. The model (RISEv2) was initialized via supervised learning on human lichess crazyhouse games and it was evaluated after 50th model updates during reinforcement learning.
- initial model: RISEv2-mobile.zip
CrazyAra 0.7.0 (Model-OS-45) played Multi-Variant-Stockfish-x86_64-modern 2019-12-03 in 100 matches (TC 15min+10s).
Eight examples of these games have been included in the master thesis
- Deep Reinforcement Learning for Crazyhouse by Johannes Czech.
50 random opening starting positions have been used out of a list of 81 openings. The positions have been provided by FM Optilink:
Hardware Setup
- Memory (RAM): 31,4 GiB
- Processor - AMD® Ryzen 7 1700 eight-core processor × 16
- Graphics - GeForce GTX 1080 Ti/PCIe/SSE2
- OS type - Ubuntu 18.04.3 LTS, 64-bit
UCI-Options
Multi-Variant-Stockfish (2019-12-03)
All default except:
- Threads: 8
- Hash: 4096
CrazyAra 0.7.0
- Search_Type: MCTS
- Context: gpu
- Device_ID: 0
- Batch_Size: 16
- Threads: 2
- Centi_CPuct_Init: 250
- CPuct_Base: 19652
- Centi_Dirichlet_Epsilon: 0
- Centi_Dirichlet_Alpha: 20
- Centi_U_Init: 100
- Centi_U_Min: 100
- U_Base: 1965
- Centi_U_Init_Divisor: 100
- Centi_Q_Value_Weight: 0
- Centi_Q_Thresh_Init: 50
- Centi_Q_Thresh_Max: 90
- Q_Thresh_Base: 1965
- Max_Search_Depth: 99
- Centi_Temperature: 80
- Temperature_Moves: 0
- Centi_Temperature_Decay: 92
- Centi_Node_Temperature: 200
- Virtual_Loss: 3
- Nodes: 15000000
- Allow_Early_Stopping: True
- Use_Raw_Network: False
- Enhance_Checks: False
- Enhance_Captures: False
- Use_Transposition_Table: True
- Use_TensorRT: True
- Model_Directory: model_os_45/
- Move_Overhead: 50
- Centi_Random_Move_Factor: 0
Nodes per Second (NPS)
- Multi-Variant-Stockfish: 6.7 Million NPS
- CrazyAra 0.7.0: 7-10 K NPS
Result
Score of CrazyAra-0.7.0-Model-OS-45-batch-16-temp-200 vs stockfish-x86_64-modern 2019-12-03: 61 - 36 - 3 [0.625]
Elo difference: 88.7 +/- 70.2, LOS: 99.4 %, DrawRatio: 3.0 %
100 of 100 games finished.
All games can be downloaded here: