serin-delaunay / tracery_game Goto Github PK

Yes please!

More than 1 start state should be available, and the resulting bot should be able to offer any of them randomly as a start tweet.

This requires changes to the grammar generation and game classes.

Doesn't affect end-user experience, but it'll be a whole lot nicer to look at.

Current scheme is direct base 62 on game state. With slightly larger state spaces densely explored this will inevitably include slurs and other potentially triggering words.

Also, base 62 requires capitals, which slow down manual typing of codes and introduce ambiguity between I and l. A smaller base would allow these problems to be resolved, and could also limit possible damage.

Instead of direct state->code conversion, use a space of codes slightly larger than the explored space, choose from them pseudorandomly, and retry when unwanted words are encountered.

This may not necessarily be practical (especially for Twine that includes JS), but it would be awesome!

Just because text display didn't work out for spacing reasons, doesn't mean we can't make the noughts and crosses bot screen reader-friendly.

[code] [input] doesn't really work well for testing in the CBDQ/CMTS interface because Chrome's double-click-to-select functionality is inconsistent about including or excluding spaces.

An improvement would be to allow arbitrary strings in between, delimited by word breaks. This would include commas and spaces, though some pathological replies (with input strings before the active input string) may be interpreted in an unintuitive way.

In noughts and crosses, the AI currently selects a random good move at compilation time.

It should instead explore all good moves, and choose a random result when sending a reply.

This requires changes in the AI routine, the the generation of the reply specification, and the state tree exploration.

This game: https://en.wikipedia.org/wiki/Sprouts_(game)

Game length, and hence state graph size, is dependent on number of initial points, so with a small enough number of points it should fit in 4MB. The question is whether that number of points is large enough to allow an interesting game.

Options: 9‚7‚2‚4‚5‚1‚8‚3‚6 is not a good look.

It would be good to be able to put this on PyPI. Needs some normalisation of package structure first.

The AI in noughts and crosses should have a small chance of making a mistake (choosing a move from a uniform distribution). This requires #1, and to minimise file sizes requires the replies to be placed in two list variables (#minmax# and #random#). The actual reply is chosen by a biased selector (approximately 99:1?).

In noughts and crosses, the game-over display needs to use custom text rather than inherit from the xo library. This is a good location to actually use Tracery grammars.

Options and code display should also probably be removed.

Comment all the things.

This one: https://en.wikipedia.org/wiki/Sim_(pencil_game). According to Slany (1999), the state space is small enough.

Primary question for me is user interface. How to communicate clearly in a toot image the numbering of all 15 edges in the graph K_6?

In the noughts and crosses example, AI move evaluations are performed by https://github.com/dwayne/xo-python. This library evaluates each board independently, which is inefficient when trying to map the whole state graph with appropriate decisions. In #24, evaluation time is almost up to 3 minutes. This can be improved by using a different AI scheme where decisions are saved and reused.

A direct port of Oregon Trail may be impractical due to state space size, but some simplifications should allow a reasonably faithful adaptation. We could even SVGify the graphics!

When inputs a and b both lead to the same game state or distribution of states, they should both be detected by a single regex. this will reduce the size of the replies file.

• Cache state map and AI moves in files in order to generate grammars more quickly (#32)
• Alt-text
• Highlight triangles that cause loss
• Record vertex permutations and report them in replies
• Allow input formats more general than 1 2, eg. 12, 1, 2, 2 1
• Empty lines should be drawn under coloured lines
• Reduce output SVG size so it fits in CBTS preview box (#34)
• State strings should be compressed (#33, #6)
• Needs messages for game over states
• AI's optimal_play key should be expanded to classify all options as optimal, suboptimal, or losing, for random selection
• Messages/"taunts" during game to indicate whether the AI expects to win or lose with perfect play. Could even be dependent on previous state and determined by reply selection

The circle example was broken by features added in order to accommodate noughts and crosses. It needs to be updated.