Quote:Original post by DaAnde
I feel that an AI should be the same as a regular player so in my picture above the AI needs to find some tinder in order to start a fire.
Well, based on your example image I am going to assume that you have a simple grid-based map with three possible cell types : tree, tinder, and open space.
Typically, in exploration type problems you first want to identify elements of the map that are relevant to your particular search goal, and make logical assumptions based on these elements as humans tend to do in real life.
If finding tinder is the goal, then logically we would look for tinder in the near vicinity of wooded areas as opposed to open-areas. Given the one example map you created, this is the only assumption that can be derived from the map that could aid in finding the timber more quickly.
To represent the assumption that tinder would more likely be near trees, you would weight the cells adjacent to tree cells in the grid with more favorable weightings, with these rewards diminishing the farther away from the tree you get. Your agent would then commence a walk through the grid, making greedy decisions and choosing successor cells with higher values where available. Doing this would make the agent prioritize searching wooded areas over searching open-space. Previously-visited cells would have their weights diminished to represent the fact that they have already been explored and are no longer of significant value to the search.
The proper term for these weighted grids would be "influence maps" I believe. There are plenty of papers available on the net discussing influence maps, so you should be able to devise a pretty workable solution to this problem using influence map techniques.