Hi guys,
I'm a researcher at University of Alberta working with Michael Buro. I've recently posted about the type of research I am interested in, and how Spring can be involved. Details in this thread: http://spring.clan-sy.com/phpbb/viewtop ... =1&t=13867
Those outline my long-term plans. Our short-term plan is to develop an abstract fast-forward simulation automated planner for AIs to use to assess the outcome of their actions in C++. It is without a doubt an ambitious project, but some work has been done already previously (see paper links in previous thread and Game AI Wisdom 4) and we may have some undergrad students working with us on the project.
The library would work on abstract levels, meaning it would not necessarily care about the granular details of the RTS game implementation, but rather the details in abstract space. What I mean by that is that this library would convert any situation to an abstract form, do all of its reasoning and simulation in the abstracted world, and return some information back to the caller. The reasoning is done by simulating the game into the future and seeing what could happen. This can be done quickly in abstract space, and the hope is that the abstraction is good enough that the information gained from these simulations can be used to make decisions.
We do plan to design an evaluation function as well so that these simulations can stop early once the evaluation function predicts a loss or win with high probability. This will be done later, but would ultimately allow the library to be used in real-time.
All this said, we do hope to implement this and use it for the upcoming RTS Game AI Competition, http://www.cs.ualberta.ca/~mburo/
I'm posting here because we may be interested in trying to get Spring to use librtsplan as well as a proof of concept. ORTS and Spring are quite different, so having a library which can be used by both projects for their AIs would be a great thing. AI developers for Spring might be interested in using this if it works well, and I may need help figuring out how the AIs work in Spring in the near future.
librtsplan
Moderators: hoijui, Moderators
Re: librtsplan
Um, may I refer you to P.U.R.E.? It's a small enough game design that it might be a suitable research environment for testing your conceptual framework. Overmind was structured to be 100% compatible with existing AIs, whereas Resistance is going to be structured around the new possibilities given to players by LUA, while attempting to be compatible with existing AIs to some extent.
I'm interested in this approach to the problems of making an AI behave "intelligently", although I suspect that your main problem is going to lie in the area of determining a valid context for the abstract modeling.
I.E., you'll run into problems such as the A.I. determining that it cannot use a MortarTank and a collection of Heavy Shells and Sniper Shells to assault a given defense, because the weighting of probable outcomes within the sim of the sim is unlikely to take into consideration the proper tactical deployment on the game board. On the other hand, it may assume that such a tactical positioning is a given, and the tactical logic may fail to accomplish the task.
The idea that you can provide at least a basis for seperating the two, at least for conflict-outcome forecasting, is a very interesting concept, however. I don't think it has a lot of applicability within the context of a non-cheating AI, but in a cheating AI, that has full access to the current position of units on the map, it would probably work quite well. It'd almost certainly work best if humans could provide the AI with a control of some sort- a config file or what-have-you- that gave the strategic-level forecaster a better idea of what real outcomes to expect from a given unit's tactical deployment.
I'm interested in this approach to the problems of making an AI behave "intelligently", although I suspect that your main problem is going to lie in the area of determining a valid context for the abstract modeling.
I.E., you'll run into problems such as the A.I. determining that it cannot use a MortarTank and a collection of Heavy Shells and Sniper Shells to assault a given defense, because the weighting of probable outcomes within the sim of the sim is unlikely to take into consideration the proper tactical deployment on the game board. On the other hand, it may assume that such a tactical positioning is a given, and the tactical logic may fail to accomplish the task.
The idea that you can provide at least a basis for seperating the two, at least for conflict-outcome forecasting, is a very interesting concept, however. I don't think it has a lot of applicability within the context of a non-cheating AI, but in a cheating AI, that has full access to the current position of units on the map, it would probably work quite well. It'd almost certainly work best if humans could provide the AI with a control of some sort- a config file or what-have-you- that gave the strategic-level forecaster a better idea of what real outcomes to expect from a given unit's tactical deployment.
Re: librtsplan
I like the idea... having read an article about emergent systems just yesterday (http://www.theatlantic.com/doc/200204/rauch - 4 years old, but still good) I think you could be onto something.
Re: librtsplan
Interesting, please keep us noted on your efforts.
Re: librtsplan
Will do. We hope to do a lot of work on it in the next 2 months, mostly using ORTS AI agents as test mechanisms. If we can in the time we have, we would like to create a Spring AI semi-concurrently to show that similar benefits can be achieved in Spring. When I come to learning how AIs work in Spring, I will surely have a lot of questions.
I expect that the feedback from the Spring community will be a big help in the development of this library.
I expect that the feedback from the Spring community will be a big help in the development of this library.