Iterative Functional Reactive Programming With The Nu Game Engine

1) The document describes an iterative functional reactive programming (IFRP) approach used in the Nu game engine that is a hybrid between classic imperative game programming and first/higher-order FRP systems. 2) IFRP uses pure functions to derive successive game states from user-defined events in a classically-iterative tick-based style rather than using continuous or discrete time-parameterized functions. 3) The IFRP approach provides multiple levels of abstraction, including observations to treat events as first-class collections, chains to allow procedural expression across events, and declarative forwarding of simulant field values.

Uploaded by

Dragos Corlet

Available Formats

Download as PDF, TXT or read online on Scribd

0% found this document useful (0 votes)

207 views

Iterative Functional Reactive Programming With The Nu Game Engine

Uploaded by

Dragos Corlet

Available Formats

Download as PDF, TXT or read online on Scribd

You are on page 1/ 4

Iterative Functional Reactive Programming

with the Nu Game Engine

An Informal Experience Report

NOTE: This is an entirely informal paper, originally targeted for a simple post on a forum of
http://lambda-the-ultimate.org, but due to my inability to achieve a readable formatting with the tools
of said forum, is made available via this PDF.

So, I have this purely functional game engine in F# - https://github.com/bryanedds/FPWorks

For readers who wish to build some familiarity with the system before reading this admittedly terse
description, please look here for Nu's high-level documentation https://github.com/bryanedds/FPWorks/blob/master/Nu/Documentation/Nu%20Game%20Engine.pdf?
raw=true
The Nu Game Engine certainly qualifies as 'functional reactive' in that 1) it is reactive in that it uses user-defined events to derive successive game states.
2) it is functional in that is uses pure functions everywhere, even in the event system.
However, I cannot describe it as the typical first-order or higher-order FRP system since it uses neither
continuous nor discrete functions explicitly parameterized with time. Instead it uses a classicallyiterative approach to advancing game states that is akin to the imperative tick-based style, but
implemented with pure functions.
Thus, I gave it the name of 'Iterative FRP' (or IFRP for short). It's a purely-functional half- step between
classic imperative game programming and first / higher-order FRP systems. There are plusses and
minuses to using this 'iterative' FRP style In the plus column 1) It's perfectly straight-forward to implement most any arbitrary game behavior without worrying
about things like space or time leaks, or having to invoke higher-ordered forms of expression.
2) End-users don't need to understand new forms of expression beyond primitive functional
expressions to encode their game logic. No lifting is required (though it is available).
3) It has a level of pluggability, data-drivenness, and serializability that are extremely difficult to
pull off in FO/HOFRP systems.
4) None of its design is the subject of open research questions, and is therefore simple enough for
non-PhD's to understand as well as complete in its implementation.
In the minus column 1) Debuggability suffers from the declarative nature of the API. For example, in Visual Studio while
stalled on a breakpoint, we cannot inspect the value of a simulants property or field simply by
mouse-hovering over its reference (or in this engines equivalent, its proxy). In order to inspect a
simulant, we must enter something akin to the following into the Watch window
Debug.Entity.peek(entity, world), ac
The first thing to notice here is that the peek function is called with a C#-style syntax, which
beyond just having to type this out rather than hover, is inconvenient. Also, we may optionally
suffix the expression with ,ac in order to have the results of the debug query update as we step
through the code.
It would be nice if I could contrive of some sort of debugging plugin to make this experience
easier, and perhaps even somehow find a way to re-enable the evaluation and visualization of
the simulant s properties / fields via a mouse hover. However, I presume both, and especially
the latter, to be quite involved if not infeasible.

2) Raw performance is good, but is not, in will never likely be, best in class. Lets face the fact that
many game programmers are speed addicts, and while their endless thirst may never be
satisfied, the engine has low-level imperative optimizations that make it more than fast enough
when used with care.
3) Verbosity is an issue also due to the indirect nature of accessing simulant properties / fields. As
mentioned above, one cannot access simulant fields with a simple property access expression,
but must be accessed by calling a function that takes a world value as a parameter. Dually,
outside of the Chain monad (which will be discussed later), the world value must be manually
thread through most engine operations.
Neither of these is a big deal, but they might be off-putting people accustomed to the
expedience of imperative programming.
Despite these minuses, I currently conclude IFRP is appropriate for such a general-purpose game engine
because
1) From my experience, games, and game technology in general, refuse to conform to any single
system of thought, often demanding certain features be implemented with escape-hatch
approaches like encapsulated mutation in order to be efficient, or with lower-level forms of
functional expression for flexibility.
2) Just because the default means of expression is not as abstract as that of FO/HOFRP systems, it
does not mean there are no available ways to 'climb into' more abstract forms of expression.
To expand on point 2, here are some of the higher forms of expression provided by the API Here we have an expression form called an 'observation' that allows combinations and transformations
of events like so let observation =
Simulants.Gameplay |>
observe (Events.Click ->- Simulants.HudHalt) |>
sum Events.Update |>
until (Events.Deselect ->- Simulants.Gameplay)

This affords us the ability to treat events like first-class collections.

Here we have an expression form called a 'chain'. A chain is a monad that allows a procedural-style
expression to span 0 or more events while also taking the world as an implicit state let chain = chain {
do! update ^ character.SetActivityState ^ Action newActionDescriptor
do! during (ActivityState.isActing << character.GetActivityState) ^ chain {
do! update ^ fun world ->
let actionDescriptor =
match character.GetActivityState world with
| Action actionDescriptor -> actionDescriptor
| _ -> failwithumf ()
let world = updateCharacterByAction actionDescriptor character world
runCharacterReaction actionDescriptor character world
do! pass }}

We also have the ability to compose chains overs observations like so -

let observation = character |> observe Events.Update |> until (Events.Deselect ->- Simulants.Gameplay)
runAssumingCascade chain observation world |> snd

Finally, there is a means to declaratively forward the changing value of a simulant's field to that of
another let world =
world |>
(bob, bob.GetVisible) *-> (jim, jim.SetVisible) |>
(jim, jim.GetVisible) /-> (bob, not >> bob.SetVisible)

Here, *-> denotes the forwarding of the value of Bob's Visible field to Jim's Visible field. To throw a
monkey-wrench into the declaration, /-> is used to in turn forward the value of Jim's Visible field back to
Bob's Visible field, albeit not'd and with cycle-breaking so that the circularity of this expression is broken
appropriately. This is a contrived example, but illustrative of the API's expressiveness.
So, depending on the nature of the game behavior you want to implement, the API provides enough
surface area to do things at different points on a spectrum of flexibility and abstraction. But there is a
down-side to an API with such a larger surface area... With multiple expression forms and levels of
abstractions at which to operate, the learning curve can be steepened. Additionally, imperative
programmers may be bothered by the performance compromises implied by pure functional
programming compared to doing everything in-place, even if performance is more than satisfactory for
most games.
Compared to the more declarative FRP styles, I believe that this iterative approach of FRP is inherently
more complete, understandable, scalable, and flexible than the more declarative FRP models (FirstOrder, Higher-Order, Arrowized), and that these properties of completeness, understandability, et al,
are the minimum necessary for developing commercial games. And as to how this approach compares
to imperative / OOP style of game development I can only hope that this approach, or at least some
other functional approach, will eventually come to replace the current methods. Game development
has, in my mind, grown far too complex to make current approaches satisfactory.