Post by alandau on Mar 18, 2015 18:06:20 GMT -8
I just added a couple of bug fixes to GGP-Base.
The first fixes how the ProverStateMachine handles recursive rules. The only game "in the wild" that I know of that's affected is amazonsTorus; the old implementation would give non-deterministic results on that game, but the new one should always be correct. This is the last correctness bug in the ProverStateMachine that I was aware of.
The second affects the Relationizer, which is used in the OptimizingPropNetFactory. This is a component designed to find sentence types in games that look state-dependent, but actually maintain the same values throughout the game. An example is the walls in ghostMaze2p. (This is a trick to make writing visualizations easier; visualizations have access to the game state, but not the rest of the game description. By adding this information to game states, the visualization doesn't need to have these elements hard-coded.) It then rewrites the game so these are constants ("relations"), not part of the game state.
The Relationizer bug was that it wasn't correctly handling these sentences if their init sentences were defined by rules, rather than listed explicitly. I'm doing a pass of adding bases and inputs to a number of games, and in these cases it makes sense to have the init and base sentences both depend on a separate constant type. The OPNF will likely fail on these games if the bug fix is not applied.
The first fixes how the ProverStateMachine handles recursive rules. The only game "in the wild" that I know of that's affected is amazonsTorus; the old implementation would give non-deterministic results on that game, but the new one should always be correct. This is the last correctness bug in the ProverStateMachine that I was aware of.
The second affects the Relationizer, which is used in the OptimizingPropNetFactory. This is a component designed to find sentence types in games that look state-dependent, but actually maintain the same values throughout the game. An example is the walls in ghostMaze2p. (This is a trick to make writing visualizations easier; visualizations have access to the game state, but not the rest of the game description. By adding this information to game states, the visualization doesn't need to have these elements hard-coded.) It then rewrites the game so these are constants ("relations"), not part of the game state.
The Relationizer bug was that it wasn't correctly handling these sentences if their init sentences were defined by rules, rather than listed explicitly. I'm doing a pass of adding bases and inputs to a number of games, and in these cases it makes sense to have the init and base sentences both depend on a separate constant type. The OPNF will likely fail on these games if the bug fix is not applied.
