Re: [Freeciv-Dev] [PATCH] Agent calls are never equal if they are to different agents
On Wed, Jun 17, 2009 at 5:35 AM, Madeline Bookmadeline.b...@gmail.com wrote: For cma in particular: it is slow (no CPU computation should take longer than a second; the algorithm does not scale), inefficient (too much client-server communication making it unwieldy in online games), does not adapt well to non-default game rules (the ugly cross-dependence of common/ files and untested behavior for other rulesets), and fails to accomodate even basic game situations that occur again and again (no way to prevent greedy tile grabbing, no way to fix tile usage, recomputation right at the very moment when client needs to be responsive, etc., etc.). I'm not sure that the _start_ of the turn is the right time to run the CMA. So many times I've forgotten to check all my cities after founding my 14th, and the next turn all hell breaks loose, because the CMA never had the opportunity to fix the civ-size unhappiness. I'd rather pretend I didn't just consider multithreading. I'm not too fussed about problems with other agents though; really only with agents-in-general. (I mention in passing that the entire attribute system should be removed, and probably would have been long ago, were it not for the single annoying dependence of the cma on it.) Would you suggest that client-side code should rather keep state in a section file? (Or even in .civclientrc itself.) If all goes to plan I'll want to remember things like The Georgians have stabbed me in the back 7 times, but the Sami have always been there to help me. Things that cannot, even in principle, be calculated from the game state. Every potential agent implementation must consider whether it fails in the above mentioned respects at least. Oh, my agent will almost certainly fail here. Epic fail. In general there is also the question whether client side automation afforded by agents is even desirable for freeciv, as then it may degenerate in to battle of computer programs rather than players. FWIW I'm working for the machines to hasten this battle of computer programs to defeat the humans. I just want to see how credible a client-side AI I can make. (Apropos, I seem to recall someone else having gone this way, but I can't find the announcement. Anybody else remember who / better search terms?) But I do understand: as a human player I also dislike the idea of twitch-like AI reaction times. For example in the case of the city worker twiddling problem that cma tries to alleviate (player does not have to micro-manage citizens, in theory), one should also consider fixing the design of the game itself. That is, instead of throwing AI at the game problem, consider finding and proposing better game mechanics that would not suffer the same player annoyances. Some of the proposals I've seen would also make it easier for an AI to accurately value a site for building a city. I'm finding it hard to analyze even how I choose places myself. Then there is the current problem of agent implementations almost always requiring to keep track of some form of state between activations, but the design of freeciv packets makes this quite cumbersome at best (cf. the request id hacks used by cma). In brief, there is poor support for stateful client side programming (other parts of freeciv suffer from this too, and I have a solution in mind, but have not yet had time to make a test implementation). This reminds me of another obstacle to easy agent programming: by the time the agent gets its callback, the entity causing the event might be gone. This manifested in a segfault when my unit-changed callback tried to print a killed unit's name. What is the intellectual status quo in freeciv regarding reference counting things like units, cities (tiles don't die)? I'd like it if the client-side unit/city objects hung around a little bit longer, at least until the agents are done using them. I can also foresee problems in trying to correlate consequences with actions. My agent will eventually have to know that a particular unit-changed callback relates to a dsend_packet_unit_move() that it issued earlier. Finally I would just like to mention lua as a much better candidate for work than the agents framework. Extending and implementing all non-resource intensive code in a lua engine would do wonders for AI programming, both on the server and the client side. At least so my dream goes, assuming that the lua integration is done right. ;) I like the idea of a scripting language hooked in, but for my code I think it might suck. I have to compensate for my Artificial Stupidity with zillions of CPU cycles. If my AS ever has to wait for a human player, it isn't doing enough number crunching! Thanks for your comprehensive reply. ___ Freeciv-dev mailing list Freeciv-dev@gna.org https://mail.gna.org/listinfo/freeciv-dev
Re: [Freeciv-Dev] [PATCH] Agent calls are never equal if they are to different agents
Just to throw in a couple cents here... I don't fundamentally disagree with what Per is saying here, but I don't think that it's all gloom and doom. Take AI government choice for example. To make the 'correct' choice of a government, a server-side AI can actually change the government, recompute all the cities and check against other choices. A client-side AI can't do that, or at least can't do that in a way which loads down the server (and a client-side AI is what we're talking about here). The CMA on the other hand can load down the server (if I recall; it's been awhile) because the client can keep asking for cma_results from the server to see if they're ok to do. Perhaps throttling these requests is a solution. I wholeheartedly agree that generalized rulesets have made it hard on AI, and doing a lot of the heavy-lifting pregame, after the rulesets have been received by the client, but before the game begins is probably the only way to go without your supercomputer. Despite this, humans can do pretty well with the amount of game state that the server sends their clients, so I expect that client-side AI can do the same with just as much information and more attention to detail. -mike On Wed, Jun 17, 2009 at 06:55:08PM +0200, Per Inge Mathisen wrote: Just to echo what was said earlier - the idea of client side agents is fundamentally a bad one because it requires too much information to be kept and sent to the clients. Only on the server can you access the amount of game state required for such agents without running head first into race conditions and then deadlocks when try to fix that because data has to be distributed over the network link and goes out of sync. It took CMA a very long time and a lot of work to reach its present state of stability. And CMA only does city tile management. I have also changed my mind about the desirability of a computational game AI. The amounts of CPU time required to compute best outcomes increases in such a dramatic fashion when the game rules are generalized, that it is simply not worth it. Good rules of thumb that can be written as AI scripts specifically for each ruleset will be faster, easier to write and maintain, and (therefore) probably even lead to better outcomes. This was not an easy conclusion for me to reach, because I spent a lot of time on the current computational AI. Hindsight is nice. If you want to throw computations at Freeciv AI problems, I recommend writing a program that analyses a ruleset then generating script rules that can be used later for a game AI. That way you can use as many CPU cycles as you want without anyone complaining. - Per ___ Freeciv-dev mailing list Freeciv-dev@gna.org https://mail.gna.org/listinfo/freeciv-dev ___ Freeciv-dev mailing list Freeciv-dev@gna.org https://mail.gna.org/listinfo/freeciv-dev
Re: [Freeciv-Dev] [PATCH] Agent calls are never equal if they are to different agents
On 15/06/2009, Bernd Jendrissek bernd.jendris...@gmail.com wrote: On Tue, Jun 9, 2009 at 3:44 AM, Madeline Bookmadeline.b...@gmail.com wrote: You should know that the agents framework suffers from a number of design problems and has not been actively maintained by anyone in a long time, the original author no longer begin involved in freeciv development and nobody else really interested in learning how it should work or how to fix it and make use of it. Could you elaborate please? I haven't been subscribed in a few years so I'm not up-to-date with current memes. You don't need to do my homework for me; I'm just looking for *what* to look for. Design problems worries me, if I'm going to be depending on agent callbacks. For cma in particular: it is slow (no CPU computation should take longer than a second; the algorithm does not scale), inefficient (too much client-server communication making it unwieldy in online games), does not adapt well to non-default game rules (the ugly cross-dependence of common/ files and untested behavior for other rulesets), and fails to accomodate even basic game situations that occur again and again (no way to prevent greedy tile grabbing, no way to fix tile usage, recomputation right at the very moment when client needs to be responsive, etc., etc.). (I mention in passing that the entire attribute system should be removed, and probably would have been long ago, were it not for the single annoying dependence of the cma on it.) Every potential agent implementation must consider whether it fails in the above mentioned respects at least. In general there is also the question whether client side automation afforded by agents is even desirable for freeciv, as then it may degenerate in to battle of computer programs rather than players. For example in the case of the city worker twiddling problem that cma tries to alleviate (player does not have to micro-manage citizens, in theory), one should also consider fixing the design of the game itself. That is, instead of throwing AI at the game problem, consider finding and proposing better game mechanics that would not suffer the same player annoyances. Then there is the current problem of agent implementations almost always requiring to keep track of some form of state between activations, but the design of freeciv packets makes this quite cumbersome at best (cf. the request id hacks used by cma). In brief, there is poor support for stateful client side programming (other parts of freeciv suffer from this too, and I have a solution in mind, but have not yet had time to make a test implementation). Finally I would just like to mention lua as a much better candidate for work than the agents framework. Extending and implementing all non-resource intensive code in a lua engine would do wonders for AI programming, both on the server and the client side. At least so my dream goes, assuming that the lua integration is done right. ;) All that said, if you are some programming wizard and have already made some significant improvements to the agents system and implemented some useful features, feel free to post your patches for possible inclusion in the development version; at least we would have something more fruitful and concrete to discuss rather than past mistakes and vague generalities. 何かがモニターのピクセルを一つずつ食べてしまう。 ___ Freeciv-dev mailing list Freeciv-dev@gna.org https://mail.gna.org/listinfo/freeciv-dev
Re: [Freeciv-Dev] [PATCH] Agent calls are never equal if they are to different agents
On Tue, Jun 9, 2009 at 3:44 AM, Madeline Bookmadeline.b...@gmail.com wrote: You should know that the agents framework suffers from a number of design problems and has not been actively maintained by anyone in a long time, the original author no longer begin involved in freeciv development and nobody else really interested in learning how it should work or how to fix it and make use of it. Could you elaborate please? I haven't been subscribed in a few years so I'm not up-to-date with current memes. You don't need to do my homework for me; I'm just looking for *what* to look for. Design problems worries me, if I'm going to be depending on agent callbacks. ___ Freeciv-dev mailing list Freeciv-dev@gna.org https://mail.gna.org/listinfo/freeciv-dev
Re: [Freeciv-Dev] [PATCH] Agent calls are never equal if they are to different agents
On 08/06/2009, Bernd Jendrissek bernd.jendris...@gmail.com wrote: While trying to add an agent, I've found this patch necessary. This is against 2.1.9, as I have nothing closer to HEAD available. Your patch looks alright, as far as I can tell and as far as I know about the intended behaviour of calls_are_equal() (should it care if the agent serving the calls is different?). Well, except for false which should be FALSE. You should know that the agents framework suffers from a number of design problems and has not been actively maintained by anyone in a long time, the original author no longer begin involved in freeciv development and nobody else really interested in learning how it should work or how to fix it and make use of it. BTW, does anyone maintain a (read-only) git mirror? SVN just isn't an option for me. I seem to be the only maintainer using git, and I would second a notion to setup a git mirror. Maybe I will at repo.or.cz, if I feel like it later. In the meantime I just use git-svn and I suggest you do the same (I know it is kind of a pain to setup). By the way, submit your patches to the new bug tracker: https://gna.org/bugs/?func=additemgroup=freeciv 私はガス状惑星で生まれたんだよ。 ___ Freeciv-dev mailing list Freeciv-dev@gna.org https://mail.gna.org/listinfo/freeciv-dev
