Prometheus: Overview

Last time I gave a brief introduction to my understanding of the agent-oriented programming (AOP) paradigm. To me, it’s a deeply interesting perspective from which to consider the operation of a computer program: as though it were a human system composed of a number of people interacting and working towards their own desires to achieve an overall system goal. At a high level, designing such a system is akin to designing a business process and selecting humans to fill the roles.

I imagine it’s something Alan Turing would have considered inevitable. If computers could be intelligent to the extent that they were conversationally indistinguishable from humans, then surely we could slot them into our logical processes in place of such a human.

Of course, a good little software engineer shouldn’t jump right into solutions; engineering is about repeatable processes, and the traditional object-oriented processes are not entirely appropriate for the level of abstraction that AOP offers.

Prometheus is a design methodology developed specifically for agent-oriented systems. It is similar in the sequence of activities to the more traditional object-oriented methodology, providing a number of concepts represented in a number of diagrams and detailed descriptions which compose a complete design. It goes through three familiar phases of specification (essentially requirements), architectural design and detailed design, with a number of artefacts that are developed concurrently and iteratively. Many of these artefacts can be cross-checked against each other to ensure consistency along the way, and the Prometheus Design Tool helps to automate this auditing.

The overall methodology is well described in the 2004 paper The Prometheus Methodology, and represented in the diagram below, with the phases as horizontal swimlanes and the artefacts as boxes in the appropriate phase. The solid arrows show how each artefact is derived from others, and the dashed arrows represent the cross-checks you can perform between them.

Prometheus overview

A high-level overview of the Prometheus methodology.

It’s a neat little diagram, and provides a good visualisation of the way in which various artefacts are developed and feed into other artefacts and so on. As a high-level diagram it doesn’t capture a lot of the detail (which the paper goes on to explain), so for a little more precision I prefer to break the phases down into separate, more comprehensive diagrams.

I’m hoping to find some time to cover each of them in future posts, but it’s a hefty process and we’re still busy trying to fit everything together outselves. I’ll try to divide them and conquer each stage in a separate post, then combine them at the end with a practical Prometheus tutorial, with reference to some of the work that has gone into our year-long software project.

As an addendum, Prometheus was created by Lin Padgham and Michael Winikoff, researchers from RMIT University’s Agents Group, just down the road from my University of Melbourne. For the record, Melbourne also developed an AOP design methodology called ROADMAP, but it’s much more abstract and incomplete beyond the requirements phase. Padgham and Winikoff wrote a book on Prometheus called Developing Intelligent Agent Systems: A Practical Guide if you’re interested in learning more, and there are a number of papers (that say much the same thing) with a wealth of references at the end of each.

* Without diving too deep into semantics, a process is a series of activities performed to achieve an end, whereas a methodology is a combined set of principles, methods and processes; a process is a subset of a methodology. Prometheus provides a set of tools as well as a way to go about using them, and hence is considered a methodology. This kind of completeness is essential for something so fundamental to provide real value.

Agent-Oriented Programming: A Brief Introduction

Not quite the same, but much catchier.

Agent-oriented programming (AOP, not to be confused with aspect-oriented programming) is a programming paradigm in the same way that object-oriented programming (OOP) is; it provides a set of concepts, and a way to think about the world in terms of those concepts. AOP is a more recent development, and still an area of considerable research and standardisation. Wikipedia traces OOP back to the 1960s, while AOP came about from research into artificial intelligence by one Yoav Shoham in the 1990s. As someone who is fascinated by new ways of thinking about software and its development, AOP is of great interest to me, and I’ve been quite fortunate to fall into a year-long university project centred around it (on the Android platform, appropriately enough).

So what is AOP? What is an agent (more precisely a “software agent“), and how does one orient one’s programming around them? AOP sits one level of abstraction above OOP, such that agents are effectively abstractions of objects. Intuitively, agents can be thought of as software entities with intelligence, just as we think of science fiction AI robots. They are machines, or in this case computer constructs, but at runtime we provide them with instructions as if they were people, like everyday colleagues. In particular, we try to avoid micromanagement of communications between agents; ideally, we would like to provide them with a task to perform and allow them to talk with each other to determine how best to accomplish it.

As if computers weren’t already making humans redundant in enough ways, taken to the logical extreme you might think of AOP as a blueprint for the obsolescence of human labour. Agent-oriented software engineering (AOSE) goes through the typical software engineering phases of requirements and design, but it does so in, again, a more human fashion. It’s more similar to designing a business process workflow than a set of modules; the focus is on the dynamic communications, the flow of information through the system, rather than the static structure, and the communication takes the form of messages passed between ‘roles’. Roles are more or less what you would expect: responsibility for an area of functionality, and the separation of concerns principle applies to keep individual roles coherent and independent. These roles are eventually adopted by agents, and then each agent has its capabilities specified hierarchically (multiple levels of sub-capabilities to achieve each capability), and the abstractions are peeled away layer by layer until each agent is composed of a set of concise, almost atomic, functions. Ultimately quite similar to OOP, but overall we’re more concerned with what an agent can do, and which other agents might be interested in that service.

The main application of AOP that comes to mind is artificial intelligence for autonomous robots, and the canonical example is one such robot that inhabits some arbitrary environment which it scans for rubbish, and upon detecting some rubbish, picks it up and drops it off in a rubbish bin. Even in such a simple case (well, it’s no driverless car), it’s interesting to consider what types of components are required. One common class of agents have beliefs, desires and intentions (hence BDI agents), just like humans, and can operate autonomously with very few ‘moving parts’:

  • Beliefs represent what the agent knows to be true or false about its environment (e.g. there is some rubbish at location (X, Y)). Beliefs are updated when the agent perceives some unexpected change in the environment (much like an event-driven programming model);
  • Desires describe the agent’s goals, or the state it would like the environment to be in (e.g. no rubbish at any location (X, Y));
  • Intentions are the actions which the agent can take to affect its environment, based on its beliefs, in order to change the state of the environment to achieve its desires (e.g. if some rubbish exists, the agent will intend to pick it up; having picked it up the agent will intend to move to the rubbish bin, and then drop it off, and so on).

Firstly, it’s interesting to consider that these basic constructs alone can theoretically describe much of the field of physical human endeavour (where the environment is the physical space-time continuum that we inhabit). However it may also be interesting to consider what human activities cannot be modelled in such a way. For example, I can’t quite picture how machines will be able to emulate the mental creativity of humans (though I’m sure they will eventually), and this is difficult to model since there is no tangible environment and often no precise desires. Empathy, or the natural understanding and sharing of emotion, would also be difficult to emulate for similar reasons, but also because an emotionless machine wouldn’t have that individual experience to fall back on. It comes down to the idea of understanding, which raises the question of what is meant by the phrase machine learning. I consider this roughly equivalent to becoming statistically more accurate by the modelling of some logical process, but whether this is actually increasing any kind of understanding on the part of the machine really depends on your definition.

Of course, there are many less fluffy applications for agent-oriented programming, and I look forward to exploring some of them in future posts. It’s a genuinely thought-provoking topic, and I hope I’ve got your mind ticking over too. It’s a field that appears easy to understand (they’re just like people, right?) though very difficult to really get your head right around, but one that I think might just be worth the time.

Entity/Component Game Design That Works: The Artemis Framework

Artemis logo

In my previous post, Entity/Component Game Design: A Primer », I reviewed some of the main differences between traditional object-oriented game design and the emerging paradigm of entity/component game design. I deliberately avoided talking about Artemis, and focused instead on the weaknesses of the alternative approaches, because I believe Artemis is able to resolve many of those problems.

Here I want to talk about the Artemis framework for a while. I keep saying I’ll write about it, but I wanted to get a little more familiar with it first. Well, I’ve been playing with it pretty solidly for about three weeks now, and I feel like I’ve got enough experience to form some kind of an opinion. Although there’s a brief overview on it’s homepage, I want to present my own perspective of it, and talk about some of my reasons for falling in love with it.

Big thanks to the authors Arni Arent (appel) and Tiago Costa (Spiegel) for all their hard work and support!

The Artemis approach

The Artemis framework inherits the concept of Entities representing bags of Components, but it does away with the OO concept of encapsulating behaviour within the components. Instead, it adopts a data-driven design involving the separation of the data and logic of each component.

In essence, the Components serve as little more than bags of data. They have no update function, no internal logic, just getter and setter functions to expose their current values.

The logic is instead encapsulated in the Systems that accompany the entities and components. A ‘System’ is simply an object that reads and updates the data in any relevant components. You could say it’s simply the update function of the relevant components refactored into its own object.

I say “relevant components” because of how the systems do their processing. Instead of the game iterating imperatively through all of the entities and updating each in turn, the systems take more of a functional programming approach. Each system specifies a set of components that it’s interested in, and then in each frame it processes only those entities that contain all specified components.

Furthermore, each system exists to perform a specific role, updating all instances of the relevant components before passing control to the next system. You could say that the components are processed in cross-sections across all entities at once, rather than entities (and all of their child components) being processed in chunks.

Advantages

That’s more or less how Artemis goes about its business, but without having used it, it may not be immediately obvious why this is a Good Thing. Following are a few reasons why I believe it’s a Very Good Thing. Of course, you can only truly appreciate the details after working with it, so I’ll try to keep the discussion at a high level.

Communication

The magic of systems is that they have no reservations in accessing multiple types of component at a time. The systems compose the communication between related components so that the components themselves don’t need to. Systems typically process one entity of interest at a time, and for the period in which they’re processing, they can combine any of the components within each entity however they like, with no structural side-effects.

For example, a MovementSystem could process all entities with Position and Velocity components, and each frame it would update the values of each Position component given the associated Velocity component. A RenderSystem could process all entities with Position and RenderableMesh components, and it would draw each mesh to the screen given the values from the updated Position component. In neither case do the components require any knowledge of each other.

Reusability

Refactoring the logic this way essentially gives you two layers of components – data components (Components) and logic components (Systems) which read/write the data components. Since the Components are pure data, they should never need to change. A 3D Velocity vector will always have X, Y and Z components, and so will be perfectly reusable in any 3D scenario.

The logic in the systems will depend on the rules of the game world, and they can be interchanged as needed. Since each system is typically self-sufficient enough to be autonomous, changing the behaviour of a system will not necessitate a change in the behaviour of any other system.

Say you want/have to change your game from a 3D free-roaming game to a 3D side-scroller. For side-scrolling movement, all you need to do is plug in a side-scrolling MovementSystem (it could be exactly the same, but simply discard the X component of the velocity), and it will automatically apply to all entities with Position and Velocity components. No components or other systems need to change as a result of the change of MovementSystem, since no components or systems even know it exists. (You may then need to update the CameraSystem and ControlSystem as well in this example, but the same process applies.)

Not prescriptive

Because the entities are simple bags of components, and the components are simple bags of data, as a developer you really only need to concern yourself with crafting the systems. The systems can easily take care of composing the relevant components within each entity, and beyond that they have no restrictions in how they process things. This means you can focus on constructing effective systems to solve problems in creative ways.

For example, I have a single CollisionSystem which detects and records all collisions in a single pass, after which I have an OO hierarchy of collision handling classes down which collisions are passed to be handled appropriately. I have an asynchronous TimerSystem which, together with a Timer component and external handler class, lets me attach and listen for delayed events with registered callbacks on individual entities with a single static function call. This may not be the “right” way to do it (and there often is no “right” way), but that’s up to me as the developer. I have the freedom to experiment, and to choose what works best, without sacrificing any part of the framework.

Convenience

All Artemis processing occurs within a managed World object. For convenience, the world object contains four managers that make your job almost too easy:

  • EntityManager manages the creation and removal of entities and their components, and allows components to be retrieved from entities by type.
  • GroupManager allows entities to be assigned to unique groups (string IDs, such as “ENEMIES”) and retrieved as a group at any time in any system.
  • TagManager allows individual entities to be assigned uniquely identifying tags (again, string IDs, such as “PLAYER”) and retrieved individually, similarly to groups.
  • SystemManager stores all of your systems, allowing them to be retrieved individually by type if necessary (for instance, the RenderSystem may need to communicate with the CameraSystem).

Performance

The Artemis library itself is very lightweight (for some sense of comparison, Artemis is 30KB compared to the ~1.5MB of Slick + LWJGL), and the internals are built for performance at larger scales. I haven’t yet reached the point where performance becomes and issue, so I can’t comment on quite how far that is, but that alone is an endorsement. Author testing has managed to create/remove a few thousand rendered entities per second, a number which goes up to millions without rendering.

On top of that, the functional programming approach to component processing has inherent potential for performance gains, if taken advantage of. I haven’t read too much into this yet as it hasn’t been necessary, but I’ll talk about it at a later date.

Conclusion

Despite appearances, this is just a static outline of how the Artemis framework fits together and some of the things it does well by itself. I’m not going to dive into all the different things Artemis allows you to do; I’ll try to cover some of those in separate posts in the near future.

I will however say that when using Artemis, the development process changes radically (for the better). You can tackle the implementation of one aspect of your game at a time in isolation. You have a clear implementation path to follow. As your library of components grows, adding new entity types to the game is as simple as creating a new Entity object and adding a different set of Components (or just modifying parameters), and they are automatically handled by the relevant Systems that you already have in place. At that point, it’s almost trivial to read in all your entities from a text file and be truly data-driven.

It’s a pleasure to work with, and every day I’m thinking of new ways to do things that make it even easier. If this article piqued any curiosity or excitement in your inquisitive mind, then I strongly encourage you to investigate using Artemis with Slick. Start by downloading the demo game, StarWarrior, and browsing through the source code. Then try making some small changes to transform it into a different game; perhaps something like Artemoids? Where you take it next is up to you and your imagination.

Just remember to keep an open mind, and have fun! I’m no expert, but I’m more than happy to give whatever advice I can, and the Slick forums are always friendly.

Resources

  • Artemis Entity System Framework – The official homepage of the Artemis project. It’s quite new (about a month old), and rather small as I mentioned. From the homepage you can get the compiled JAR, source (JAR, SVN, Git), JavaDoc, and the demo StarWarrior game (inspired by Space Invaders). There’s also an unofficial C# port going on.
  • Artemis thread in the Slick forums – Artemis was originally developed for Slick, and so that’s where most of the discussion is so far. I should mention however that it’s not dependent on Slick, and appel is already using it with libgdx.
    • For those of you interested in what I’ve been discussing, I joined in on page 12.
  • Entity Systems are the future of MMOG development (T=Machine) – A very thorough application to MMOs, and a primary inspiration for Artemis.
  • Anything on this site with the “artemis” tag – The count of this content will be increasing consistently in the future, because Artemis is a lot of fun to talk about.

Entity/Component Game Design: A Primer

Components of Crysis

Crysis rendered as bags of components (click for detail).

I want to talk a bit about the Artemis framework I’ve been playing with, but it won’t mean much without some background knowledge of the entity/component paradigm of game design. This post, then, is a quick review of my experience with entities and components, and how they relate to the more traditional object-oriented (OO) paradigm, and in particular, what sucks about both.

If you’re familiar with the entity/component paradigm already, you may just want to skip over to my detailing of Entity/Component Game Design That Works: The Artemis Framework » and read through this at your leisure. There’s a few connections, but I’m trying to keep it at a pretty high level for now.

Object-oriented games

In a traditional OO design, every “thing” in the game, be it a tank, a bullet, a light or a sound effect, is defined by its own class. Naturally there is a great deal of similarity between some classes, and so they can inherit common behaviour from more abstract classes. For example, a Tank and and a Car are quite similar: they both drive around, are affected by physics, have some visual form, and make sounds occasionally. This behaviour is common to all vehicles, and so can be inherited from an abstract Vehicle class.

This is quite intuitive, relatively effective, and allows the programmer to add a Truck by inheriting that same behaviour from Vehicle and assigning it a truck model. Of course, both Vehicles and Projectiles are affected by physics, so both abstract classes can inherit the properties and behaviour of a more abstract PhysicalObject, and so on.

The design of an object-oriented game will require a detailed design of this class hierarchy, with each concrete in-game object inheriting from a tree of increasingly abstract classes, right up to something like GameObject, which may just represent anything with an ID assigned to it. If this hierarchy is planned out well before implementation, then it’s quite possible to build a large, complex game leveraging OO inheritance. After all, that’s how most games are built.

However, the deeper your class hierarchy grows, the more brittle and fragile it becomes. If (read: when) the requirements change after implementation begins, reflecting those changes in the code will often require adding and/or removing functionality to some of those abstract classes. Such changes will flow on to affect all subclasses, whether or not those subclasses need the changes. The result is messy code additions that get pushed up towards the root of the (upside-down) tree in order to provide the changes to a greater number of subclasses. (The alternative is duplicating and encapsulating the changes in each concrete class that requires them, which is arguably worse still.)

Entity/component games

One of the core principles of good software design is modularity: breaking down the core components of a system into individual, independent units, and then building the system as a whole by plugging these components together. There are many benefits of modularity, including, most relevantly, flexibility. If one component needs to change, a modified component can be plugged in to replace the old component without affecting the rest of the system. So long as the rest of the system can talk to the new component in the same way as they did the old one (i.e. it conforms to the same interface), then they are oblivious that a change even occurred.

Since sometime around the turn of the millennium, the idea of breaking down the traditional game class hierarchy into a more modular design has started to spread. It’s not a particularly new idea, it has no obvious source, and now it seems to be everywhere in every conceivable form if you go digging for it. I want to focus however on the ‘traditional’ entity/component model, and later contrast it to that adopted by Artemis.

Traditional components

My first encounter with a (relatively) formal definition of entities and components in a game setting was in Game Programming Gems 6, where it was discussed as a “Game Object Component System”.

Here’s the deal. Forget about your class hierarchy, throw out your painstakingly-drawn inheritance trees, and take off your OO hat for a minute. With entities and components, we’re only interested in concrete “things” – Tanks and Cars, not abstract ‘Vehicles’. By the time you’re done, any game object hierarchies you have should be almost completely flat.

Some terminology: an Entity represents a concrete “thing” in the world, such as a Tank. However an Entity has no Tank-specific logic; indeed, an Entity barely has any logic at all, and is little more than an ID.

The real magic comes from the components it contains. A Component is a module of functionality, call it an attribute if you will. Components are things that Entities have, things like Position, Velocity, RenderableMesh, and PhysicalBody. An Entity is little more than a bag of Components; it is quite literally the sum of its parts. Importantly, the Entity has no explicit knowledge of what parts it contains, meaning all entities can be treated the same way by the rest of the game.

This is possible because the components take care of themselves, regardless of which entity they belong to. For example, the RenderableMesh component contains the capability to render a 3D model. The model is assigned to the component, and the component is assigned to the entity, so the entity doesn’t need to know what the model looks like. All the entity needs to do is call some generic update function on each of its components each frame, and each component will do its own thing. A RenderableMesh, for instance, could draw itself to the display.

The beautiful advantage of such a system is that the components are generic, they perform a single role, and they perform it the same way regardless of their parent entity. As such, the RenderableMesh of a Tank object would draw itself the same way as that of a Car object, the only difference being the shape of the mesh assigned to each component. As a result, different types of entities can be manufactured easily by plugging different reusable components into an empty entity.

This is great news for maintaining flexibility during and after development! Changes to entities will typically involve changing one or two components in isolation, without changing any unrelated components or polluting other entities. New functionality can be added with the independent addition of new components.

Weaknesses

The unfortunate weakness of this system is that there are inherent relationships between the components which requires them to be coupled in some way or other. They are, after all, parts of a single whole. For instance, the RenderableMesh knows how to render, but it doesn’t know where to render without consulting the Position component. The Velocity component can very well keep track of velocity, but it’s not much good without being able to update the Position component.

One proposed solution is to push this ‘shared data’ up into the Entity itself, so that all components can read and write to the position variable. However this leads to an analogous problem to that of the traditional OO hierarchy; it’s a slippery slope. Once you push up position, you may realise that movement also needs to know how much health the entity has, and so you can either communicate with the Health component or push health up into the entity. And hey, rendering and collision handling need to know the physical dimensions as well. And so on.

Another solution is to allow components to hold references to each other and communicate directly, but this has multiple drawbacks associated with it. For one, it couples the components very tightly, making them difficult to reuse under different circumstances. In addition, all these references need to be assigned when the entity is created, but without the entity’s direct intervention, which is a non-trivial problem in itself. One possible sub-solution is to implement some elaborate system of runtime inter-component dependency resolution and injection, which adds a mountain of complexity to what should be a molehill of an obstacle (simply plugging in components).

An alternative to holding direct references between components is to attach a message dispatching system to each entity, allowing components to fire events when something interesting happens, and handle other events of interest fired by other components. This has the advantage of decoupling the components which is nice, but comes at the cost of, once again, increased complexity and a persistent performance penalty.

Even if you manage to get all your components hooked up and working together, completely encapsulating the logic related to each component within the component itself is fraught with the danger of your components becoming bloated with functionality. In theory, a Physics component should handle all physical interaction between its parent entity and the rest of the world, but does the knowledge of the rest of the world really belong in a component? Perhaps we can split out the physics calculations into a centralised physics manager… So then which components have logic and which don’t, and how much?

The answer?

There are many, many different approaches to working around these multifarious concerns, some of which I mentioned above, but they all seem to raise more problems of their own. Does there exist some architectural solution to circumvent these obstacles, not by climbing over them but by simply changing lanes? I’m a strong believer that any software design is possible with enough ingenuity, and I’ve found the Artemis framework to be pretty ingenious. Read on for why I think Artemis represents Entity/Component Game Design That Works ».

Resources

There are buckets of great articles out there explaining the entity/component game design paradigm from different perspectives, and it’s the kind of concept where your understanding really benefits from a broad exposure. Below is just a handful of the ones I’ve found useful in my adventures. You can, of course, find plenty more by Googling “entity component game”. Please mention any other particularly useful ones you’ve found in the comments and I’ll add them in too!

Jonathan Blow on How to Program Independent Games

Braid image

I’ve just got done listening to a university presentation by Jonathan Blow, creator of the fantastic Braid, on his insights into independent game programming. It’s very good.

The talk is about 38 minutes with another 40 of Q&A, and he talks mostly about the “aesthetic” he maintains for being “brutally productive” the way he is required to be to create (and finish) something like Braid (with around 90,000 lines of code) by himself in just a few years.

Since it’s a talk to a group of computer science students it’s a little technical but not especially so. The key points are about not making things more complex than they need to be; effectively Keep It Simple, Stupid.

  • Avoid premature optimisation, including sophisticated algorithms and data structures, which are essentially about optimisation.
  • Don’t generalise a program more than necessary, even to the point of avoiding refactoring a 1,000 line function.
  • A “good” programmer will get things done quickly and solidly, and simply, with an emphasis on actuallygetting things done.

This is very congruent with my mentality, which is why you’re unlikely to see my writing about fancy algorithms and data structures (or the dreaded topic of “web scale“; warning, justified language). I like doing things, creating things that don’t already exist, and although I’m still relatively inexperienced I do very much hope to one day reach that point where I can put together something like Braid given a good enough idea.

Anyway, I’ve embedded the talk below (audio + slides) because I can (the embed code was provided). As an added bonus, you get to listen to him talk about The Witness (or more precisely, it’s editor) which he’s happily showing off on a different screen.