As the uni semester brings further deadlines, my progress slows immensely.
There are some problems to address that need to be fixed by 25th March as given by my proposed timetable. Though these were covered in the last post, I’m readdressing them:
- The biggest problem at the moment is I-pieces being put in terrible positions. Code needs to modify it so I-pieces are always put vertical if no horizontal available and should be put next to climbs to minimise jaggedness of terrain.
- Placing pieces in the lowest ideal position. this can be addressed by multiplying the idealness of a slightly worse position by 2. Will need to debug it.
- Placing pieces so they lay as horizontal as possible if there are no better positions.
- Top-of-the-field rotation restrictions. When a chasm is directly below an I-piece for example, it will want to rotate to suit it and drop. However, the top of the field stops it from doing so and causes the agent to think the piece is rotated, thus netting a terrible horizontal I-piece reward. May not apply for just I-pieces.
- Finding the piece at the start of a new episode.
- Not really a problem, but needs to be done by the 25th. Eligibility traces. Shouldn’t be too tricky. The functions are there, just gotta put the list/stack/thing in.
I believe that’s everything. Something else to note is debugging output. Perhaps have a line of code that sends to agent_message to toggle debugging output. When on, displays all the necessary information (perhaps the set of substates when calculated, the piece location at every step to make sure it matches up, the goal position, etc).