Closed thebot002 closed 2 hours ago
Implemented 16th of September Fixed 21st of September
Illustration:
Results: The performance is much better and now reached 100% of convergence when testing on an environment issue from a probability equal to the trained one.
Now the model needs to be tested on different source positions to verify it also performs good with the source being anywhere, also in places it was not trained on.
Model with macro-tiling 5x5 and micro-tiling unknown
Convergence 100%
Path:
Note: We see that the agent, when reaching the center performs exploration which is good.
Performances across different macro-tiling and micro-tilings:
A specified in #50 The source-tile is divided into a subgrid of its own.
Technical detail: The observation probabilities in any of the sub-tiles is equal to the observation probability in the macro-tile that would be in their place. This is to not give any more information as necessary to the agent.