The stage1-mcts dataset contains 8,001,990 evaluated positions (datapoints) generated by the Monte-Carlo Tree Search (MCTS) with random playouts. The average number of MCTS simulations per datapoint is: 1,359,479. Table below shows more details.
| group | num datapoints | average num simulations |
|---|---|---|
| nsim=1mln | 7,049,309 | 1,000,000 |
| nsim>1mln | 952,681 | 4,019,422 |
| all | 8,001,990 | 1,359,479 |
Some positions have been evaluated with over 1 mln simulations to improve the quality and diversity of the dataset.
Datapoints are grouped by their depth, which is defined as the number of moves (plies) for a given position. The histogram below shows the number of datapoints per depth in the stage1-mcts dataset.
MCTS datapoints were created using: ../scripts/mcts_evaluate.py, ../scripts/mcts_generate.py, ../utttcpp/src/mcts-evaluate.cpp and ../utttcpp/src/mcts-generate.cpp. Initial depth for self-playing was progressively deepened to avoid repeating computation.
MCTS datapoint format:
evaluatedState{UTTT_STATE NUM_VISITS NUM_WINS NUM_DRAWS NUM_LOSSES} evaluatedActions{SYMBOL INDEX NUM_VISITS NUM_WINS NUM_DRAWS NUM_LOSSES, ...}
download link: https://drive.google.com/file/d/1y7V1HWKdykuxD6Q8PNhydjPrHxkBr6vu/view?usp=sharing (815 MB, after extracting: 2.6 GB)
The stage2-nmcts dataset contains 8,023,020 evaluated positions (datapoints) generated by the Neural Monte-Carlo Tree Search (NMCTS) with Policy-Value Network trained on the stage1-mcts dataset. The average number of NMCTS simulations per datapoint is: 10,000. Table below shows more details.
| group | num datapoints | average num simulations |
|---|---|---|
| nsim=10k | 8,022,643 | 10,000 |
| nsim>10k | 377 | 20,610 |
| all | 8,023,020 | 10,000 |
Datapoints are grouped by their depth, which is defined as the number of moves (plies) for a given position. The histogram below shows the number of datapoints per depth in the stage2-nmcts dataset.
NMCTS datapoints were created using: ../scripts/nmcts_evaluate.py, ../scripts/nmcts_evaluate_parallel.py, ../scripts/nmcts_generate.py, ../scripts/nmcts_generate_parallel.py, ../utttcpp/src/nmcts-evaluate.cpp and ../utttcpp/src/nmcts-generate.cpp. Initial depth for self-playing was progressively deepened to avoid repeating computation.
NMCTS datapoint format:
evaluatedState{UTTT_STATE VISIT_COUNT STATE_VALUE_MEAN} evaluatedActions{SYMBOL INDEX VISIT_COUNT STATE_VALUE_MEAN, ...}
download link: https://drive.google.com/file/d/1BxsJ8BmcluRLI6WK8ElIVVwLhCtjn2kQ/view?usp=sharing (481 MB, after extracting: 2.1 GB)