Designing Naturalistic Simulations for Evolving AGI Species

Christian Hahm

Department of Computer & Information Sciences, Temple University, Philadelphia, PA, U.S.A.

Keywords:

Adaptation, Artiﬁcial Evolution, Genetic Algorithm, Virtual Environment, Artiﬁcial Life, AI Embodiment,

Evolutionary Simulation.

Abstract:

This paper identiﬁes basic principles for designing and creating evolutionary simulations in the context of

general-purpose AI (AGI). It is argued that evolutionary simulations which employ certain nature-inspired

principles can be used to evolve increasingly intelligent AGI species. AGI frameworks are particularly suited

for evolutionary experiments involving embodiment since they can operate arbitrary evolved bodies. Once a

designer manually deﬁnes a simulation’s initial conditions, each run is an automated exploration of a novel

subset of species. In this way, naturalistic simulations generate huge amounts of empirical data for evaluating

the robustness of AGI frameworks, along with many promising species that can be later instantiated in other

simulated environments or even physical robots for practical applications.

1 INTRODUCTION

Virtual simulation goes naturally with AI: since both

are implemented in digital computers, it is almost triv-

ial to interface the two. We should consider this a

huge beneﬁt, since virtual environments (which, for

this discussion, can be referred to using other words

e.g., a “world”, “universe”, “reality”, “simulation”

etc.) are highly customizable by the designer. To per-

form “simulation” in the context of AI means:

1.) To design and create a virtual world (in a com-

puter, using a “simulator” software), and

2.) To embody one or more AI agents in that

world.

The purpose of simulation is to provide AI agents

a controlled environment in which to interact and

learn from their simulated experience, such that we

may study the agents (e.g., to observe their processes

under certain conditions) or utilize them in some real-

world application once they are sufﬁciently trained.

An agent is considered “embodied” in an envi-

ronment so long as it has sensorimotor mechanisms

(Wang, 2009), meaning sensors (i.e., tools on the

body which transduce environmental signals as inputs

to the mind) and actuators (i.e., tools on the body

controlled by output signals from the mind) in that

environment. The agent receives sensory signals to

monitor the environment, and can send output sig-

nals to move its body. Like a “brain-in-a-vat”, an AI

system treats sensory signals identically regardless of

whether they come from simulated virtual sensors or

real-world physical sensors.

Artiﬁcial general intelligence (AGI) is a term

which can be used to delineate cognitive-inspired

computer systems with full autonomy and general-

purpose ability. The possible applications of simu-

lation to AGI research are numerous, but mostly in-

volve optimizing the system on either the object level

or the meta level. In object-level testing, we embody

some agents and let them learn simulated tasks au-

tonomously. In meta-level testing, we vary conﬁg-

urable cognitive parameters to test various “personal-

ities” of AGI individuals on object-level tests. Cog-

nitive parameters may range from scalar control val-

ues (e.g., decision thresholds, attention decay, etc.) to

structures in the system’s architecture (e.g., buffers,

memory, etc.). The exact nature of cognitive parame-

ters depends on the selected framework.

Since AGI systems are general-purpose, they

should be capable of operating arbitrary bodies. In

considering embodiment, we can expand the notion

of parameters from cognitive values to include phys-

ical structures like the system’s morphology, sensors,

actuators, and physiology (bodily parameters). Any

parts of the system can be parameterized, except those

core aspects which we desire to remain static. When

aspects of the system’s mind and body are parame-

terized, it evokes the possibility of creating varied AI

“species”: agents which share a common core frame-

work of intelligence, but vary by their exact cognitive

296

Hahm, C.

Designing Naturalistic Simulations for Evolving AGI Species.

DOI: 10.5220/0011299600003274

In Proceedings of the 12th International Conference on Simulation and Modeling Methodologies, Technologies and Applications (SIMULTECH 2022), pages 296-303

ISBN: 978-989-758-578-4; ISSN: 2184-2841

processes, bodily form, and sensorimotor capabilities.

Evolutionary simulations automate the process of

ﬁnding and optimizing such mind-body parameteri-

zations. Traditionally, evolutionary simulations prop-

agate species that maximize a quantitative “ﬁtness”

function. Natural evolution admits no such ﬁtness

function, and to simulate it requires a slightly dif-

ferent treatment than traditional genetic algorithms.

Instead of optimizing genomes to maximize a pre-

deﬁned ﬁtness function, “naturalistic simulations” re-

quire agents to prove their reproductive ﬁtness in-

dependently, encouraging the evolution of intelligent

autonomy. If the selected AGI framework is ﬂexible

and the simulated environment is sufﬁciently rich, the

resulting species should exhibit interesting adaptive

behaviors.

2 RELATED WORKS

Evolving autonomous machines is not a novel con-

cept; one of the earliest and most famous attempts at

elaborating this idea is mathematician John von Neu-

mann’s self-reproducing automaton. He hypothesized

a baseline level of complexity that allows the evolu-

tion of increasingly complex systems (von Neumann,

1966, p.78-80), realized by the three sub-processes

of reproduction and evolution: duplication, modiﬁ-

cation, and instantiation of genomes. In his context,

the automaton system itself performs all three sub-

processes, including modifying the genome.

Though “AGI” is often construed as human-level

intelligence, our simulated agents will be more like

primitive animals than humans. They will not be ca-

pable of manually modifying their own genes; the

task of genomic modiﬁcation belongs to the simula-

tion. As long as the agents handle reproduction au-

tonomously, the simulation design will handle the ac-

tual evolutionary processes behind the scenes. Game

engines such as Unity3D (Juliani et al., 2018) are the

ideal simulator software since they facilitate building

simulated physical worlds and arbitrary scripting.

(Holland, 1992) formalizes a methodology for ap-

proaching so-called “problems of adaptation”. This

work led to the popularization of genetic algorithms,

an evolutionary computing method by which to ex-

plore a space of evolving structures using concep-

tions of ﬁtness, genetics, and reproduction. Genetic

algorithms are arguably the best approach for simu-

lating natural evolution, because they implement the

general principles that make natural genetics adap-

tive. Namely, genetic algorithms use “survival of the

ﬁttest”, recombination, and random mutation to ex-

plore a certain space of evolving structures in search

of more optimal (i.e., “ﬁt”) structures.

(Sims, 1994a; Sims, 1994b) use Holland’s genetic

algorithms to evolve various agent bodies and their

neural networks, including by facing the agents off

in direct physical competition (Sims, 1994a). The

agents are encoded using a highly compact and ﬂexi-

ble genetic language in the form of directed network

graphs. Physical body parts contain sensors (e.g.,

contact sensor, light sensor) or actuators (e.g., rota-

tional force on joints), and each body part has a neural

circuit allowing for some level of distributed control.

(Soros and Stanley, 2014) propose four neces-

sary conditions for open-ended artiﬁcial evolution:

1&2.) new and existing individuals must be required

to meet some non-trivial minimum criterion (MC) be-

fore they can reproduce, so species do not degener-

ate into trivial-behavior automatons, 3.) individuals

must autonomously meet that MC, 4.) the complex-

ity of the individual’s genotype can grow unrestricted.

The naturalistic MC is the ability to reproduce, in-

cluding the requisite apparatus and motivation (Soros

and Stanley, 2014, p.2-3).

(Stranneg

ard et al., 2020) establishes a “generic

animats” framework in which simulated organisms

are deﬁned by homeostatic variables, sensors, pat-

terns, motor, actions, and reﬂexes. (Stranneg

ard et al.,

2021) formalizes ecosystems containing inanimate

objects and organisms that have unique properties

(e.g., digital “animals” have energy levels, environ-

mental objects have certain chemical reactivity) and

common properties (e.g., all objects have mass).

For this discussion we will assume Wang’s work-

ing deﬁnition of intelligence: “the capacity of a

system to adapt to its environment with insufﬁcient

knowledge and resources” (Wang, 1995; Wang, 2019,

p.13; p.17-18). According to this deﬁnition, the min-

imum requirement for intelligence is to operate un-

der an Assumption of Insufﬁcient Knowledge and Re-

sources (AIKR). A system operating under AIKR: 1.)

uses constant computational resources, 2.) works in

real-time, and 3.) is open to new tasks regardless of

their content. In plainer terms, such a system is like a

living organism: it is a fast, ﬁnite agent that operates

effectively even under uncertainty, while constantly

learning from its experience so as to better achieve its

goals. Cognitive frameworks designed with AIKR in

mind may be ﬂexible enough to permit the evolution

of complex adaptive behaviors.

Designing Naturalistic Simulations for Evolving AGI Species

297

3 SIMULATION DESIGN

3.1 Formalizing the Problem

Following (Holland, 1992, p.28,35), a given problem

of adaptation and the adaptive system to tackle it, in

our case evolving AI species in simulation, are de-

ﬁned by the variables: α, Ω, T , ε, χ.

The variable α = {A

, A

, ...} is the set of all

genomes that can possibly be evolved; if the simu-

lation’s genetic language permits genomes to grow

without restriction, this set is inﬁnite. An evolution-

ary run searches this set. Genome encoding should

include both cognitive parameters for the AGI system

and bodily parameters for its embodiment.

Ω = {ω

, ω

, ...} is the set of genetic operators

that can be used to modify genomes. A given op-

erator might perform some variation of crossover or

mutation.

T = {τ

, τ

, ...} is the set of possible reproduc-

tive plans, each of which is a sequence of operators

from Ω which can be used to incrementally traverse

α. One τ may be selected for the entire simulation

run, or the choice of τ could even be evolved.

ε = {E

, E

, ...} is the set of possible environ-

ments. Creating a concrete simulation requires speci-

fying the actual environment E ∈ ε.

Finally, χ represents a criterion to compare the

many possible plans in T .

So, to specify these variables is to determine the

possibilities of the evolutionary simulation. The con-

tents of plans in T depend on the operators avail-

able in Ω; for example, a plan could look like τ =

(ω

, ω

, ...). Operators in Ω depend on how

the designer speciﬁes the genetic language for α.

Therefore, this discussion will mostly ignore the de-

pendent variables Ω and T , instead focusing on gen-

eral considerations for specifying α, E, and χ.

3.2 Embodiment

3.2.1 Atoms

One major design problem of virtual simulation is that

we have to manually specify its irreducible compo-

nents. There are no interacting molecules like there

are in reality, unless we program them. We want

to simulate a physical reality efﬁciently and with ﬁ-

nite resources. We also want a ﬂexible simulation,

which requires irreducible generic components that

can combine and interact in various ways.

The simulation can implement a notion of equal

exchange, where a ﬁnite number of virtual atoms un-

derlie all structures in the simulation. There could be

Figure 1: Atoms are the underlying virtual representation of

a simulation’s structures. In this mockup, the green block

is an environmental structure consisting of atoms. Agents

could consume the atoms to exchange them for homeostatic

maintenance, physical growth, or offspring.

various types of atoms with unique properties. The

atoms are abstract intermediary components, never

explicitly seen but used justify environmental struc-

tures in E and physical expression of body parts en-

coded by α (see Figure 1). Atoms could correspond

one-to-one with size-standardized polygons or voxels.

This prevents any structure from growing dispropor-

tionately compared to the others by enforcing equal

exchange.

While pure morphology can be grown incremen-

tally, polygon by polygon, some body parts must be

hardcoded with special-purpose functionality in the

form of modality-speciﬁc sensors, actuators, and mo-

tivation. For example, vision requires a special pho-

toreceptor body part which can render a partial view-

point of the scene and communicate it to the AGI sys-

tem. One way to handle this is to exchange atoms one-

to-one for pure structure (with some distinguishing

properties) whereas require a many-to-one exchange

for structures with special functionality. For a simple

example, two red and two green atoms could corre-

spond to a single photoreceptor, meaning a parental

agent will need to consume those atoms to instantiate

an offspring with a photoreceptor part. Agents would

return their body’s absorbed atoms to E upon death

(via simulated decomposition) for recycling.

The minimum E is an environment that hosts re-

producing (thus evolving) cognitive agents and re-

spects a conservation of materials. Once this is

achieved, E can be made more complex and inter-

active using frameworks like atoms, perhaps even

including genomes for other kingdoms of life (e.g.,

plants, fungus, etc.). Various “organisms” could reor-

ganize atoms in unique ways, in a simulated “circle of

life”. Non-cognitive entities like plants are not strictly

necessary, but without them the environment will be

simplistic and should not be expected to produce very

sophisticated or interesting agents.

SIMULTECH 2022 - 12th International Conference on Simulation and Modeling Methodologies, Technologies and Applications

298

3.2.2 Sensorimotor

In order to evolve sensorimotor, ﬁrst the environment

E must simulate speciﬁc modalities (e.g., ﬂoating

odorants, sound waves), and the genomes in α should

be capable of expressing the corresponding modality-

speciﬁc sensors and actuators (e.g., olfactory organs,

visual organs, vocal cords, etc.). It is certainly worth

exploring ways to evolve arbitrary sensory modali-

ties. However, since sensors measure physical signals

(e.g., goodness of molecular ﬁt, mechanical pressure,

wavelength of light, etc.) which do not exist by de-

fault in simulation, it seems necessary to explicitly

support certain modalities in the design of E and α.

Figure 2: Sensors such as tactile sensors on a 3D model

(left, middle) should be represented “topographically”,

where physically adjacent sensors are also represented adja-

cently, such as in a network (right). (Made with Blender, blender.org).

Natural sensorimotor neurons in the body bun-

dle then connect to the brain in spatially ordered

“topographic” mappings (see (Wolfe et al., 2006,

p.64,280,397)). This helps the agent understand the

spatial distribution of its sensations. Simulated sen-

sorimotor signals could be represented in a similar

way, where sensors of the same type record their sen-

sations in a topographic map. The map should be “to-

pographic” in that sensors which are physically close

and adjacent in the simulation are also represented ad-

jacently in the map (see Figure 2).

For example, it is quite intuitive to simulate simple

vision: render a full 2D visual image of the scene, and

treat each pixel of the image as a single visual sensor

(aka photoreceptor). The situation is muddier for the

other senses. More generally, topography can be de-

scribed using a network, where edges explicitly (as in

a graph, where nearby nodes are connected) or im-

plicitly (as in a convolutional neural network, where

neighboring nodes connect to a common node) rep-

resent adjacency. A network representation might be

useful for touch perception, which relies on unevenly

distributed sensors and may not be as amenable as vi-

sion to a 2D array representation. The AGI system

can then use the topography to selectively group and

process sensations (Wang et al., 2022).

3.3 Fitness and Reproduction

Traditionally with genetic algorithms there is an addi-

tional variable, µ, specifying the function by which to

measure an individual’s ﬁtness. The simulation pro-

ceeds in discrete rounds, where after each round the

ﬁtness of individuals is quantitatively measured us-

ing µ. Then, the genetic material of a few individuals

with the highest scores are combined to create many

offspring, whereas those with lower scores are culled.

In AI research, we would ideally select a µ that mea-

sures and thus optimizes for intelligence. However,

no widely-agreed µ exists to quantitatively measure

“intelligence”, amongst humans nor across species.

Besides, evaluating agent performance with a single

number is inﬂexible, since such evaluations are more

vulnerable to “hacky solutions” and much generality

is lost as species evolve towards optimizing only one

function.

In nature, “reproductive ﬁtness” is a tautology.

There is no numeric ﬁtness measure by which na-

ture decides reproduction, instead ﬁtness is equiva-

lent to (or proven by) successful reproduction. Traits

which confer reproductive and survival advantages

in the current environment will tend to appear more

frequently in the populations than traits which pro-

vide relatively less advantage. Therefore, a naturalis-

tic simulation with an implicit µ should at the very

least yield increasingly proliﬁc species. Reproduc-

tion, as in nature, can be done sexually or asexu-

ally. Asexual reproduction is simpler to simulate and

very quickly grows the population, though agents who

are selectively sexual might induce faster speciation

and evolved capabilities, as argued in (Canino-Koning

et al., 2017).

Although such a µ frees us from needing to explic-

itly measure ﬁtness, it introduces a number of prob-

lems. First, all simulated agents need to meet the nat-

ural “minimum criteria” (Soros and Stanley, 2014):

they must be capable of reproducing autonomously.

“Capable” not only means physically able to perform

a reproductive action (as determined by the speciﬁc

simulation), but also requires the agents to be moti-

vated towards that action (i.e., to have a sex drive). In

biological contexts, bodily glands release hormones

that modulate reproductive motivation in the brain

(Wise, 1987; Fisher et al., 2006; Cummings and

Becker, 2012). A bodily origin of motivations in natu-

ral agents is interesting to notice in our context, since

it implies motivation evolves with the body.

There needs to be a tradeoff between consuming

atoms from the environment (e.g., food) and produc-

ing offspring. Consistent and balanced resource ex-

change is critical as it requires agents to earn their

Designing Naturalistic Simulations for Evolving AGI Species

299

success. The exact process of this tradeoff during re-

production depends on the simulation design and α.

One could simulate naturalistic processes like gesta-

tion (where a parent gets pregnant then gradually con-

sumes the relevant atoms to grow a baby), incuba-

tion, external egg fertilization, etc., though the bio-

logical details are not necessary to replicate, only an

energy/materials tradeoff to instantiate and grow the

offspring.

The choice of adaptive plan τ ∈ T for a given

simulation run is important, since it determines the

evolutionary trajectories through α. Yet, we do not

know the best τ to choose. χ is the quantitative cri-

terion for evaluating plans in T so as to ﬁnd the best

τ. The χ is normally something like average ﬁtness

of the population, ¯µ, so as to try various τ in appli-

cation and pick one which maximizes ¯µ, but since a

naturalistic µ is implicit, we cannot pick such a χ.

One decent option for χ is the total number of in-

dividuals produced by τ , which should result in sim-

ulations with many agents and hopefully more oppor-

tunities for beneﬁcial evolution. However, any χ or τ

can be tried — more agents by time t does not nec-

essarily mean “more-intelligent” agents by time t. It

may be possible to treat χ implicitly like µ, leaving

the choice of τ to evolution by allowing the selection

of τ to vary depending on the species (in which case,

each A ∈ α should represent its selection of τ).

3.4 Homeostasis as Seed Motivation

Figure 3: Bodily systems entering homeostatic imbalance

may trigger sensations and motivation in the mind.

Homeostasis is the tendency of a system to maintain

bodily variables within an optimal range. The purpose

of homeostasis is to keep the system alive; whenever

there is a homeostatic imbalance, the body triggers

automatic reﬂexes and signals motivation to the mind

so as to ﬁx the imbalance. (Tsitolovsky, 2015, p.3,4)

explains: “Needs are at the heart of motivations”,

and since homeostatic goals are endogenous (origi-

nating from within the agent’s own mind-body sys-

tem, rather than from an external user) they “turn [the

agent] into a subject with its own behavior”. Reﬂexes

only have very limited power to restore the body to

stability, depending on the severity and complexity

of the issue. This is why, in nature, both sensation

and motivation are signalled to the central nervous

system, so that the agent itself can restore homeosta-

sis by interacting with the environment (Craig, 2003;

Tsitolovsky, 2015).

Maintenance of the body as the agent’s source of

motivation makes perfect sense, as it keeps the agent

alive. Determining how agents source motivation is

extremely important since without motivation there is

no action. A small set of motivational signals from

the body is enough. Simple motivational “seeds” can

derive a wide range of additional motivations, causing

many complex behaviors to emerge depending on the

speciﬁc agent’s knowledge and capabilities (Hahm

et al., 2021). Evolving homeostatic systems on the

body can provide various seed motivation signals to

the agent. The more skilled the agent is at meeting

useful homeostatic survival needs, the more likely the

agent will survive to reproduce.

One could simulate homeostatic motivation

as in Figure 3, for example, within a body part

containing variable-speciﬁc sensors which activate

whenever some variable (e.g., heat, cold, pain)

veers too far from its acceptable range. The sensor

then signals both a sensation of the imbalance

S = Belief (sensation) and a negative desire to

alleviate it D = Goal(¬sensation). Alternatively,

a lack of sensation S = Belief (¬sensation) can

be paired with a positive desire to experience it

D = Goal(sensation). This model can account

for both avoidance behavior (e.g., polygons are

destroyed when they take enough damage, and

nociceptors detect damage. When the agent’s arm

mesh sustains a strong force, desire is signalled to

stop activation of the arm’s nociceptor):

{S, D} = {Belief(pain), Goal(¬pain)}

and approach behavior (e.g., green-atoms are the

agent’s store of energy, and have a property that is

detected by the mouth’s sweet-taste receptors when

eaten. When the agent’s green-atom stores drop

below critical levels, desire is signalled to activate the

mouth’s sweet-taste receptors):

{S, D} = {Belief(¬sweet), Goal(sweet)}

In the avoidance case, the agent is motivated to

stop an event (the pain, such as by withdrawing its

arm from the painful stimulus), whereas in the ap-

proach case, the agent is motivated to realize an

SIMULTECH 2022 - 12th International Conference on Simulation and Modeling Methodologies, Technologies and Applications

300

event (tasting sweet, such as by seeking and eating

green-atom). A sensation S alone is neutral informa-

tion, making the agent aware of the current situation.

However, by additionally signalling motivation D, a

homeostatic system inﬂuences the agent’s reaction to

3.5 Evolving Intelligence (Implications

of AIKR)

We have conjectured that a naturalistic simulation

should produce increasingly proliﬁc AGI species.

However, it is easy to see where this strategy will

fail. With a boxed-in environment and no energy con-

straints, agents will inﬁnitely reproduce until the en-

vironment is overﬂowing with simple reproduction-

optimizers. In the context of AGI, we want to evolve

increasingly intelligent species, though proliﬁc agents

help prevent global extinction. According to the as-

sumed working deﬁnition of intelligence, this means

we wish to improve each individual agent’s capacity

to adapt to its environment (i.e., its ability to learn

and execute skills on the object level) under AIKR,

using the evolutionary parameters we have available

(via genomic adaptation, on the meta level).

The condition of AIKR in the deﬁnition of intel-

ligence is an important nuance, because it moves us

slightly from the traditional viewpoint of intelligence

as an agent’s ability to adapt to an agent’s ability

to adapt under uncertainty and resource limitations.

So to evolve greater intelligence requires pressures in

the form of resource/energy limitations (i.e., insufﬁ-

cient resources) and uncertainty about E (i.e., insuf-

ﬁcient knowledge). If the environment E is danger-

ously uncertain but the agent manages to survive it

and reproduce, then the resultant species might not be

reproduction-optimizers in the most simplistic sense,

but rather species who are reproduction-optimizers

despite the odds. Such species will necessarily ex-

hibit intelligent behaviors that are linked inextricably

to both their mind-body forms and the environment E

in which they evolved.

In other words, since we have no ﬁtness func-

tion to pressure greater intelligence explicitly, agents

prove their superior capability in E independently by

surviving and achieving reproduction. If the environ-

ment is too gentle, there is no reason for species to

evolve better capabilities. If the environment is too

harsh, everyone goes extinct. One problem is, any ini-

tially “difﬁcult” static pressures that we design in the

environment might be overcome by some specially-

adapted species, which would then explode in popu-

lation and plateau in complexity as the evolutionary

pressure is nulliﬁed.

This means the environment needs to exhibit

adaptive, even scaling, pressures on the global popu-

lation. In this regard, the agents are their own ideal

pressure, and they must be allowed to interact and

even (implicitly) forced to compete. As one species

improves their capability in E, another will be pres-

sured to keep up or risk extinction. Macro-scale

competition between species results in some species

adapting, then others, in a neverending back-and-

forth game of genetic improvements. The result is an

“arms races between and within species” (Dawkins

and Krebs, 1979), which is precisely what is needed

to prevent intelligence from plateauing.

The uncertainty arising from an uncountable

amount of complex evolving agent-environment and

agent-agent interactions in naturalistic simulation

would overwhelm traditional AI systems. For exam-

ple, reinforcement learning agents assume states are

repeatable, and so would fail in a naturalistic simula-

tion which never exactly repeats. On the other hand,

this treatment is possible with AGI systems working

under AIKR as they are fundamentally equipped to

deal with uncertainty.

3.6 Initial Conditions

Once the atoms and encoding for α are decided, the

actual initial state of environment E must be created.

This is up to creativity of the designer, with two no-

table constraints.

Firstly, E must be contained with ﬁnite (though

renewable) resources. Consider an environment that

is just an inﬁnite 2D plane with agents and resources

on top if it. Not only would this provide agents with

inﬁnite resources, but the agents might simply spread

out and never reproduce. Those that do would have no

reason to evolve greater abilities, as they could always

migrate to a fresh new location when they use up all

the simple resources in their current location. In con-

trast, organisms which are forced together will have

no choice but to interact with each other and compete

over the limited resources, thus pressuring impressive

evolution via the arms race phenomenon.

Secondly, E must be populated with one or more

initial organisms at t = 0. The organism must at min-

imum possess reproductive ability, motivation, and

embodiment, spending energy to do work. A starting

point is the most simplistic organism possible accord-

ing to your speciﬁcation of α; perhaps a simple body

capable of movement, tactile sensors or a small eye,

asexual reproductive organ, and a simpliﬁed gastroin-

testinal system (mouth, stomach, etc.) to maintain en-

ergy levels. Such organisms should begin to ﬁll up the

space of E, and the evolutionary takeoff begins.

Designing Naturalistic Simulations for Evolving AGI Species

301

4 SUMMARY

By selecting a general-purpose AI framework and ap-

plying nature-inspired principles in simulation, we

can explore a huge variety of AGI species and even

pressure them towards greater intelligence. Naturalis-

tic simulations provide insights into the selected AGI

model and could even yield interesting or impressive

species for real-world applications.

The ﬁrst step is to deﬁne the atomic building

blocks of the simulation. These atoms provide justiﬁ-

cation in the form of equal exchange for structures in

the environment E and the agent population. Atoms

should be fairly exchanged for energy expenditure

and the volume/function of body parts instantiated

from α, such as during offspring creation. The deﬁ-

nition of “equal” in equal exchange might be arbitrary

and up to the designer, as long as the exchange rates

are coherent and invariant. Just as atom classes are

hardcoded, the simulated transduction of environmen-

tal signals to sensory signals seems to require hard-

coded functionality, depending on which modalities

the designer wishes to explicitly support.

The next step is to specify which aspects of the

AI system and body are parameterizable (and in turn,

evolvable). There are two overarching classes of pa-

rameters to evolve and encode in the genome: cog-

nitive parameters, which modify the selected AGI

model (e.g., its control system, architecture, decision

thresholds, etc.), and bodily parameters, which mod-

ify the agent’s embodiment (e.g., morphology, phys-

iology [including homeostatic systems], and sensori-

motor capabilities). As such, the selected AGI model

must expose certain cognitive parameters, whereas

the simulation designer speciﬁes the available bodily

parameters.

It is important that parameters of both types co-

evolve since their interplay may be delicate when it

comes to the agent exhibiting coordinated behavior.

Parameterization is among the most important factors

to consider when designing an evolutionary simula-

tion, since it fundamentally constrains the range of

species that can possibly be evolved. The more gen-

eral and ﬂexible these parameters are, the more op-

portunities for novel, interesting, and useful abilities

to evolve. On the other hand, too many options in

α (especially for parameters which drastically alter

the core functioning of the system) can impede the

search for high-performing species, as the simulation

may waste computational resources on offspring with

relatively incoherent mind-body parameterizations.

Motivation seeds are “externally” sourced by the

agent’s body parts during homeostatic imbalances,

though such homeostatic systems will likely only be

burdened by a species in exchange for increased sur-

vival and reproductive success. Sensation signals

keep the agent informed of its current homeostatic

outlook, while motivation signals sway the agent’s

behavior. At absolute minimum, reproductive moti-

vation should be guaranteed in every organism so as

to perpetuate evolution.

Despite hunger’s appearance as a motivation

unique to organic creatures, energy maintenance mo-

tivation (such as in the form of hunger) is likely also

essential in artiﬁcial agents, since working with ﬁnite

resources demands a work-energy tradeoff. There-

fore, an agent without hunger will quickly use all its

energy without replenishment and die before it can

reproduce. Energy limitations force an agent to be

smarter about how and when it acts so as to achieve

its goals.

In many cases it is implausible to numerically

measure performance on complex or abstract tasks.

This is especially true when it comes to measuring an

ill-deﬁned concept like “intelligence”, which has been

interpreted in various ways. A naturalistic simulation

is open-ended, rejecting any single choice for ﬁtness

function µ to evolve intelligent behavior. Instead, as

in nature, many capable species co-evolve when indi-

viduals of each species prove their own reproductive

worthiness.

Agents are pressured to evolve better adaptive ca-

pabilities when the environment contains dangerous

uncertainties (increasing the importance of operating

with insufﬁcient knowledge) and their vital resources

are strained (increasing the importance of working

with insufﬁcient resources). Relatively static environ-

mental resource pressures will become less relevant

as species evolve greater capabilities towards surviv-

ing in their environment. However, if the environment

is contained, the species themselves should exert scal-

ing adaptive pressures on each other as they evolve to

out-compete each other for the limited resources.

Though there are many open questions, natu-

ralistic simulation seems like a plausible way to

evolve AGI agents with various intelligent capabili-

ties. Species evolved in especially physically realistic

simulations could even be instantiated in real-world

robots designed analogously to the simulated bod-

ies. A simulation can be slowed down, allowing us

to observe agents interacting in real-time, or sped up,

to hasten the evolutionary process and explore new

species populations. Overall, naturalistic evolution-

ary simulations are a tool to automatically explore a

wide range of capable AGI species, collect empirical

data, and gain insights into both AI design and the

nature of “intelligence”.

SIMULTECH 2022 - 12th International Conference on Simulation and Modeling Methodologies, Technologies and Applications

302

REFERENCES

Canino-Koning, R., Keagy, J., and Ofria, C. (2017). Sex-

ual selection promotes ecological speciation in digi-

tal organisms. In ECAL 2017, the Fourteenth Euro-

pean Conference on Artiﬁcial Life, pages 84–90. MIT

Press.

Craig, A. D. (2003). A new view of pain as a homeostatic

emotion. Trends in neurosciences, 26(6):303–307.

Cummings, J. A. and Becker, J. B. (2012). Quantitative as-

sessment of female sexual motivation in the rat: Hor-

monal control of motivation. Journal of neuroscience

methods, 204(2):227–233.

Dawkins, R. and Krebs, J. R. (1979). Arms races be-

tween and within species. Proceedings of the Royal

Society of London. Series B. Biological Sciences,

205(1161):489–511.

Fisher, H. E., Aron, A., and Brown, L. L. (2006). Roman-

tic love: a mammalian brain system for mate choice.

Philosophical Transactions of the Royal Society B: Bi-

ological Sciences, 361(1476):2173–2186.

Hahm, C., Xu, B., and Wang, P. (2021). Goal generation

and management in nars. In International Confer-

ence on Artiﬁcial General Intelligence, pages 96–105.

Springer.

Holland, J. H. (1992). Adaptation in natural and artiﬁcial

systems: an introductory analysis with applications to

biology, control, and artiﬁcial intelligence. The Uni-

versity of Michigan Press.

Juliani, A., Berges, V.-P., Teng, E., Cohen, A., Harper, J.,

Elion, C., Goy, C., Gao, Y., Henry, H., Mattar, M.,

et al. (2018). Unity: A general platform for intelligent

agents. arXiv preprint arXiv:1809.02627.

Sims, K. (1994a). Evolving 3d morphology and behavior

by competition. Artiﬁcial life, 1(4):353–372.

Sims, K. (1994b). Evolving virtual creatures. In Pro-

ceedings of the 21st annual conference on Computer

graphics and interactive techniques, pages 15–22.

Soros, L. and Stanley, K. (2014). Identifying necessary con-

ditions for open-ended evolution through the artiﬁcial

life world of chromaria. In ALIFE 14: The Fourteenth

International Conference on the Synthesis and Simu-

lation of Living Systems, pages 793–800. MIT Press.

Stranneg

ard, C., Engsner, N., Ferrari, P., Glimmerfors,

H., S

odergren, M. H., Karlsson, T., Kleve, B., and

Skoglund, V. (2021). The ecosystem path to agi. In

International Conference on Artiﬁcial General Intelli-

gence, pages 269–278. Springer.

Stranneg

ard, C., Xu, W., Engsner, N., and Endler, J. A.

(2020). Combining evolution and learning in com-

putational ecosystems. Journal of Artiﬁcial General

Intelligence, 11(1):1–37.

Tsitolovsky, L. E. (2015). Endogenous generation of goals

and homeostasis. In Anticipation: Learning from the

past, pages 175–191. Springer.

von Neumann, J. (1966). Theory of Self-Reproducing Au-

tomata. University of Illinois Press. Edited and Com-

pleted by Arthur W. Burks.

Wang, P. (1995). Non-Axiomatic Reasoning System: Ex-

ploring the Essence of Intelligence. PhD thesis, Indi-

ana University.

Wang, P. (2009). Embodiment: Does a laptop have a body?

In Goertzel, B., Hitzler, P., and Hutter, M., editors,

Proceedings of the Second Conference on Artiﬁcial

General Intelligence, pages 174–179.

Wang, P. (2019). On deﬁning artiﬁcial intelligence. Journal

of Artiﬁcial General Intelligence, 10(2):1–37.

Wang, P., Hahm, C., and Hammer, P. (2022). A model of

uniﬁed perception and cognition. Frontiers in Artiﬁ-

cial Intelligence, 5.

Wise, R. A. (1987). Sensorimotor modulation and the vari-

able action pattern (vap): Toward a noncircular deﬁni-

tion of drive and motivation. Psychobiology, 15(1):7–

20.

Wolfe, J. M., Kluender, K. R., Levi, D. M., Bartoshuk,

L. M., Herz, R. S., Klatzky, R. L., Lederman, S. J.,

and Merfeld, D. (2006). Sensation & perception. Sin-

auer Sunderland, MA.

Designing Naturalistic Simulations for Evolving AGI Species

303