go-atari-montecarlo, v1.2

The library that implements an Atari Go engine based on the Monte Carlo tree search algorithm.

Remove redundant features.

Change Log

• remove:
  • separate representation of a game result;
  • node scoring by node win rate;
  • resetting of iteration terminators;
  • move searchers:
    • searcher that reuses a built tree;
    • fallback searcher that uses an additional searcher when the primary one returns an error;
• improve:
  • move selectors:
    • simplify:
      • constructing of node group;
      • architecture of move selectors;
    • optimize node scoring by the Upper Confidence Bound algorithm;
  • add a universal utility for parallel processing.

Features

• move searching via the Monte Carlo tree search algorithm:
  • move selectors:
    • random selecting;
    • selecting by a maximal node score:
      • scoring by the Upper Confidence Bound algorithm;
  • game simulating by simple random rollout;
  • tree building:
    • by a single pass;
    • by iterative passes:
      • iteration terminating:
        • by a pass;
        • by a time;
  • move searchers:
    • searcher that doesn't reuse a built tree;
• optimization via parallel move searching:
  • parallel game simulating:
    • of a single node child;
    • of all node children;
  • parallel tree building;
• easily extensible and composable architecture:
  • of move selectors:
    • of node scorers;
  • of game simulators;
  • of tree builders:
    • of iteration terminators;
  • of move searchers.

Benchmarks

Without parallelism:

BenchmarkSearch_with5Passes-8                                    300       4742169 ns/op
BenchmarkSearch_with10Passes-8                                   200       9232705 ns/op
BenchmarkSearch_with15Passes-8                                   100      13969380 ns/op
BenchmarkSearch_with20Passes-8                                   100      18659831 ns/op

With parallel game simulating of a single node child:

BenchmarkSearch_withParallelSimulatorAnd5Passes-8                100      12154846 ns/op
BenchmarkSearch_withParallelSimulatorAnd10Passes-8                50      24416774 ns/op
BenchmarkSearch_withParallelSimulatorAnd15Passes-8                30      37238231 ns/op
BenchmarkSearch_withParallelSimulatorAnd20Passes-8                30      49606338 ns/op

With parallel game simulating of all node children:

BenchmarkSearch_withParallelBulkySimulatorAnd5Passes-8            50      27779097 ns/op
BenchmarkSearch_withParallelBulkySimulatorAnd10Passes-8           20      61114056 ns/op
BenchmarkSearch_withParallelBulkySimulatorAnd15Passes-8           20     110363599 ns/op
BenchmarkSearch_withParallelBulkySimulatorAnd20Passes-8           10     183186149 ns/op

With parallel tree building:

BenchmarkSearch_withParallelBuilderAnd5Passes-8                  100      17207899 ns/op
BenchmarkSearch_withParallelBuilderAnd10Passes-8                  30      34048898 ns/op
BenchmarkSearch_withParallelBuilderAnd15Passes-8                  30      51624139 ns/op
BenchmarkSearch_withParallelBuilderAnd20Passes-8                  20      68313262 ns/op

Repository

Link: https://github.com/thewizardplusplus/go-atari-montecarlo/tree/v1.2.

Content: code.

License: MIT.

Screenshots

Comparing benchmarks of v1.1 and v1.2