Proceedings of the 18th International Conference on System Theory, SuF3.

3
Control and Computing, Sinaia, Romania, October 17-19, 2014

Information-Computation-Evolution Agent Model for

Natural Computation

Beatrice Luca
Department of Computer Science
Faculty of Automatic Control and Computer Enginnering
Iasi, Romania
Email: [email protected]

Abstract—Biological systems provide an exceptional source Similar with Richard Feynman and Stephen Wolfram [8]
of inspiration for the design of autonomous adaptive agents. views, in which the world operates on a very simple level
In this work, we propose ICE (or Information-Computation- from which its complexity emerges, we propose a description
Evolution), a new bio-inspired agent model for complex, multi- of the physical world using simple rules, with behavior corre-
agent self-organized systems, such as biological or social systems. sponding to a basic fundamental structure of information from
The model proposes a physical organization of reality using
three main components - information organization, graph-based
which complex rules emerge. Complex rules belong to more
computation and evolution. It emphasizes a strong relationship efficient fundamental structures of information (called time
between computation and the physical world out of which we and space fundamental units) and act as ‘physical laws’ in the
derive information principles and a biological description of the world. For example, complex rules can describe ‘motion’ of
world. Agents using ICE operate in an hierarchical structure objects (similar with Newton’s laws of motion) or ‘reflection’
of ‘realities’ where they have to conform to the information (similar with Newton’s law of gravity). A complex rule of
principles derived from the model and continuously explore more motion describes uniform motion or how the motion changes
complex knowledge and create new connections to learn more its direction or speed. A complex rule of reflection explains
from the world. the connections between current and future decisions using
Keywords—Hierarchical modeling, Embodiment, Classical and approximations on best future opportunities.
non-classical computation, Evolution, Information principles. Since we describe the complexity of the world emerging
from simple rules, then we imply a computable world in which
I. I NTRODUCTION objects are individuals that describe basic information using
basic rules, computation using complex rules, and other com-
From the beginning of civilization, we humans have in- plex mechanisms such as life using embodiment, evolution and
creased our understanding of reality by first decoding what cognition. The second layer of reality organizes objects into a
we see around us, and then use this information to construct family of networks representing successive configurations of
a better picture of ‘reality’ [4]. Information and computation the system. A network of objects is similar with a classical
have always been used in this process of understanding and system [9], where the results of measurements localized in
explaining the world. One approach is to view objects in terms space-time are fully determined by properties inherent in that
of information and computation and assign them complex system (its physical reality), and cannot be instantaneously
mechanisms such as life, rational and social behavior [10]. influenced by a distant event (locality). However, when objects
Some authors extend this definition of computation at different are correlated such that they can share behavior, the network of
levels of reality [5]. objects is similar with a quantum system which is fundamen-
tally non-local. That is, one is forced to choose a context of
In this work, we propose a memory evolutive system with computation (i.e., choose a goal first) in which a measurement
three layers of ‘realities’: the first layer represents the physical of an object is performed (i.e., choose a sub-goal) and then its
world, the second layer represents the memory with successive result may affect some other objects from the same context.
configurations of the system, and the third layer represents
the computable or experimental world that emerges from the The study of bio-inspired systems revitalized the study of
previous layers. Agents operating in this hierarchical structure embodiment and its implications through better exploitation
continuously explore more complex knowledge and learn using of materials, highly sophisticated sensory-motor systems and
evolutionary computation. We address the following chal- clever controls and morphological designs [7], [3], [1]. The
lenges: (a) understand and describe the physical world in terms important challenges of the field are to understand how life
of information, computation and evolution, (b) understand emerges, what are the limitations and potentials of living
embodiment and its connections to the physical world, and beings, and how its life related to body and mind? In this work,
(c) describe autonomy of agents and learning capabilities in we propose a constructive proof for the genotype-phenotype
the proposed context of evolutionary computation. To address problem that is effectively used to construct objects using
these problems, our solution takes inspiration from Newton’s their properties. Then, we describe information principles and
laws of motion and gravity to construct the physical layer, and the emergent biological view of the world. Further, we look
uses rich bio-inspired features such as embodiment, cognition at the implications of embodiment for successful design of
and evolution for the memory and experimental layers. autonomous adaptive agents.

978-1-4799-4601-3/14/$31.00 ©2014 IEEE 868

 In Section II we discuss the important features of {c1 , . . . , cNγi }, for the application of αi in W , and each
the Information-Computation-Evolution model: fundamental condition cj = {E 0 , l} checks the existence of one entity e
structures, rules and operations, objects and their properties, in- from E 0 ⊆ E, at location l ∈ M ,
formation principles, relations between objects, computational − fp is the profit function of r (i.e, fp (r)), with the default
contexts, frames and universes. In Section III we discuss the value as maximum profit that can be accomplished by r in W .
layered architecture of the model: the physical reality layer,
the memory layer and the experimental layer, and propose a The application of a rule r on entity e ∈ E in W is an
general architecture of an autonomous agent displaying self- action a and it consists of several steps:
organization and emergence. In Section IV we conclude and 1) For all application environments γi , it checks if all
discuss future work opportunities. conditions cj are true.
2) It computes a new location l0 from α and the current
II. A N EW B IO - INSPIRED M ODEL FOR NATURAL location lc of e.
C OMPUTATION 3) It moves e at location l0 .
The requirements for computational models of bio-inspired 4) It removes (or replaces) entities from W according
agents are well defined by Ingold [6]: organisms-centric biol- to a given specification (e.g., if W is a game, then
ogy is “a biology that asserts the primacy of processes over the specification could be to remove pieces that
events, of relationships over entities and of development over have been captured or to replace pieces that have
structure”. In this section we detail the main components of accomplished their goal with better pieces).
ICE - physical world and behavior, objects and their genetic In addition to their role in describing different types of
encoding of information, relations between objects, structures, behaviors, rules describe ‘motion’ and ‘reflection’ on future
and computation. opportunities. Two operators, composition (>) and reflection
(>>), inspired from Newton’s laws of motion and the gravity
A. Physical World and Behavior law, are proposed in Definition 3.
1) Rules and Operations: ICE model describes a physical Definition 3: (Operations on rules) Let e be an entity in
world with entities that generate rules, strategies as objects W and r1 = (α1 , β1 , fp1 ), and r2 = (α2 , β2 , fp2 ) are two of
physically embedded in the world, and agents that dictate what its rules. Then, the composition and reflection operators on r1
objects are used, their relations and development. and r2 are defined as follows:
Definition 1: (Physical World) The physical world is a 1) Composition: r3 is the composition of r1 and r2 if the
dynamic environment W , usually described using a map M application of r1 on e is followed by the application
with entities E that generate rules R, and agents A that use of r2 on e.
objects O to make decisions in the world: r3 = r1 > r2 , (3)
W = (M, E, R, O, A) (1) where r3 = (α3 , β3 , fp3 ) with α3 = append(α1 , α2 ),
β3 = β1 ∪ β2 and fp3 = fp1 + fp2 .
Rules are formalized in Definition 2. Each rule r incorpo-
rates a specific type of behavior α (e.g., moving left or right, 2) Reflection: r3 is the reflection of r1 and r2 if it
jumping over pieces, capturing pieces) and a set of application produces the rule r1 or r2 , with the best reflection
environments β, with practical constraints cj for the rule r10 or r20 on future opportunities. For example, r10 is
application. Practical constraints are related to the efficiency the best reflection of r1 if it follows the application
of rules and their physical limitations. A large number of of r1 on e and it produces the highest profit among
constraints represents a more complex, hence more efficient all rules applicable to e from that location.
rule, while a small number of constraints represents a physical r3 = r1 >> r2 , (4)
limitation for the rule.
where r3 = r1 if fp (r10 ) ≥ fp (r20 ), or r3 = r2 otherwise.
We make the following notations: Nα is the number of
relative coordinates in α, Nβ is the number of application
environments in β and Nγ is the number of constraints in an The composition operator describes how the motion of
application environment γ. In our model, Nβ must be equal entities change. For example, using composition we can define
with Nα because we associate one action to one application different types of motion: linear, zig-zag or circular motions.
environment. An important advantage of using composition is that composed
rules are easier to write and better understood in this form.
Definition 2: (Rules) A rule is described by its behavior
α, a set of practical constraints β, and a profit function fp : We use rk to denote the composition of a rule r, k times.
Parameter k can be static - its value is given at static time, or
r = (α, β, fp ), where (2) dynamic - its value depends on the actual state of the world
− α = (α1 , . . . , αNα ) is an ordered set of behaviors, defined and is discovered at dynamic time. Based on parameter k, we
here as an ordered set of relative coordinates αi = (∆x, ∆y)i , call complex rules static or dynamic.
where an entity is moved from the current location (x, y)c , Two examples of static and dynamic rules that use com-
− β = {γ1 , . . . , γNβ } is a set of application environments position are given in the following:
of r; each γi from β describes a set of conditions, γi =

869
 • Static rule r3 with k = 2 defined at static time: shapes the world. In the next section, we define embodiment
r3 = (r1 > r2 )2 of objects and give a simple formal proof for the construction
r3 is equivalent with (r1 > r2 ) > (r1 > r2 ). of objects using their properties.
• Dynamic rule r3 with k determined (or limited by the
physical world) at dynamic time: B. Objects and the Genetic Encoding of Information
r3 = r1k > r2 In the previous section we have seen how we can model
r3 is equivalent with r1 > · · · > r1 > r2 . the physical reality using static and dynamic rules that can
| {z }
k times share behavior. The next step is to use this reality and create
new objects that represents goals and sub-goals in the world.
The reflection operator describes motion towards the best There are two fundamental questions we address: what is an
future opportunities. Reflection rules are usually expensive object and what it is made of?
because the reflections are found at dynamic time. Fortunately,
reflection rules can become static by explicitly defining future The objects are similar with entities that already exist in
opportunities using composition with other static rules or the world, but have a higher level of manifestation. Namely,
special kind of rules called tail rule. A tail rule only has objects express goals and sub-goals using strategies and have
practical constraints and does not introduce any behavior. A properties useful for their characterization and construction.
special type of a tail rule is the empty rule, r⊥ , with no Two types of strategies are defined: goal strategies represent
behavior or constraints. goals in the world and helper strategies represent sub-goals
used at accomplishing goals. In Section II-B1, we define two
We use r[k] to denote a generalization of r which detects properties of strategies - universality and embodiment. We
best opportunities k steps ahead. show that these properties imply information principles that
Two examples of static and dynamic rules that use reflec- can be used to construct a biological image of the world. In
tion are given in the following: Section II-B2, we formalize helper and goal strategies. Finally,
in Section II-B3 we propose a solution to the genotype-
• Static rule r3 with the reflection defined at static time phenotype problem for constructing strategies.
as a single tail rule rt :
r3 = (r1 >> r2 ) > rt . 1) Theoretical Properties and Information Principles:
Given a world W , we define a new set of objects O in W
• Dynamic rule r3 with best reflections r10 and r20 as goal strategies, Sg1 , . . . , SgNg , and for each Sgi , a set of
determined at dynamic time, two steps ahead of r1 helper strategies, Shi,1 , . . . , Shi,Ni . Each goal strategy Sgi can
or r2 : be accomplished using a set of rules Ri ∈ R, which are
r3 = (r1 >> r2 )[2] . distributed among its helper strategies.
2) Fundamental Structure: So far, we have modeled the We use the notation Ng to denote the number of goal
world using static and dynamic rules. Further, we argue that strategies in W and Ni to denote the number of helper
the world has a fundamental structure of information, similar strategies for a given goal Sgi .
with bits or atoms, that represent behavior at different levels
of efficiency. Since efficiency is usually related with space Property 1: (Universality Property) Any helper strategy
and time, three types of fundamental units of information are Shi,j in W is uniquely identified by a profit function
proposed in Definition 4. fp (Shi,j ). Any goal strategy Sgi is uniquely identified by
the maximum profit of its helper strategies, i.e, fp (Sgi ) =
Definition 4: (Fundamental Units) A fundamental unit de- maxj=1,Ni (fp (Shi,j )).
scribes behavior with a given level of efficiency. Given two
parameters of the world W , pW and cW for profit and Property 2: (Embodiment Property) Any strategy in W
complexity, we propose three types of fundamental units: has an embodiment of the same information that shapes W ,
i.e., it uses rules from the same fundamental units that describe
• Basic fundamental unit (ub ) describes basic behavior. W.
ub has basic rules with profits less than or equal to
pW and complexities less than or equal to cW : It follows that these properties are useful for describing
important information principles (e.g, Evolution and Cognition
∀ri ∈ ub , fp (ri ) ≤ pW and c(ri ) ≤ cW .
Principles). From information principles a new biological view
• Time-efficient fundamental unit (ut ) describes time- of the world emerges.
efficient behavior. ut has rules with profits greater than
pW and complexities less than or equal to cW : Principle 1: (Evolution) Evolution is a high level manifes-
∀ri ∈ ut , fp (ri ) > pW and c(ri ) ≤ cW . tation of phenotypes (or strategies) derived from the following
rule of life: when strategies reproduce, their children have the
• Space-efficient fundamental unit (us ) describes com- same universality and embodiment property.
plex, space-efficient behavior. us has rules with com-
plexity greater than cW : Principle 2: (Cognition) Cognition is the ability to have an
∀ri ∈ us , c(ri ) > cW . embodiment as described by Property 2. As strategies become
more physically embedded in the world (i.e, they acquire more
The principal role of the above structure of information is rules of W ), their cognitive skills improve and increase the
to explain the origin of information and to show that embodi- chance of being used in the process of interaction with the
ment of objects is a manifestation of the same information that world.

870
 From Evolution principle it follows the neutrality of the C. Relations, Structure and Computation
genotype-phenotype map, where many mutations do no have
an appreciable effect on the phenotype. Hence, there is a We model computation based on the idea of a relation
reduced chance of becoming trapped in sub-optimal regions of digraph DS = (S, RS ) where nodes represent strategies and
arcs represent relations between strategies.
the genotype space. The process of acquiring cognitive skills
described in the Cognition principle is similar with the brain 1) Relations: We describe two types of relations, overlap-
training which rewires the brain so that it function faster and ping and goal-oriented relations, formalized in Definitions 7
more efficiently than before. and 8. Informally, two strategies have an overlapping relation if
they share the application of a set of rules, and a goal-oriented
2) Helper and Goal Strategies: Using the universality relation if one helps in the accomplishment of the other.
property, we define helper and goal strategies in Definitions
5 and 6. We use Si to denote any helper or goal strategies and
Ro and Rg to denote overlapping and goal oriented relations
Definition 5: (Helper Strategies) A helper strategy Shi,j between strategies.
is the set of all rules rk ∈ Ri with profit fp (rk ) equal to
fp (Shi,j ). The application of Shi,j represents the application Definition 7: (Overlapping relation - Ro ). An overlapping
of one of its component rules. relation is a symmetric binary relation that connects two
strategies, Si and Sj , that share rules (condition 5∗ ) and
Definition 6: (Goal Strategies) A goal strategy Sgi is the locations of application in the world (condition 5∗∗ ).
set of all rules used by its helper strategies, Shi,1 , . . . , Shi,Ni . Ro = {(Si , Sj ) : ∃Si0 ⊆ Si and Sj0 ⊆ Sj , s.t.
The application of Sgi represents the application of any of its (5)
∀ri ∈ Si0 , ∃rj ∈ Sj0 s.t. ri = rj (∗) and li0 = lj0 (∗∗)}
helper strategies.
(Si , Sj ) ∈ Ro ⇐⇒ (Sj , Si ) ∈ Ro (6)
3) The Genotype-Phenotype Problem: The genotype-
phenotype problem asks for an efficient solution to the map- Definition 8: (Goal oriented relation - Rg ). A goal-
ping of genotypes to phenotypes. In our model phenotypes oriented relation is an asymmetric binary relation that connects
are strategies with a certain appearance (or embodiment) and two strategies, Si and Sj , such that all rules in Si can be used
behavior. Their manifestation is affected by the physical world, to accomplish Sj .
development and learning. Genotypes are rules which represent Rg = {(Si , Sj ) : ∀ri ∈ Si , ri is used to acomplish Sj } (7)
the ‘genetic material’ of strategies. Throughout this work, we
address several challenging questions related to the genotype-
phenotype problem: how to map genotypes to phenotypes, (Si , Sj ) ∈ Rg =⇒ (Sj , Si ) ∈
/ Rg (8)
what are the key features of the genotype-phenotype mapping,
and how does evolution steer in the genotype-phenotype map? An overlapping relation (Si , Sj ) ∈ Ro is represented by an
arc from Si to Sj (Si overlaps on Sj ) and an arc from Sj to Si
We propose a solution for mapping complex rules to strate- (Sj overlaps on Si ). A relation (Si , Sj ) ∈ Rg is represented
gies that uses both universality and embodiment properties. by an arc from Si to Sj (Si helps in the accomplishment of
Theorem 1 provides a constructive proof which is derived Sj ). A missing arc signifies the relation is not defined or does
straight forward from the structure of the world. Similar not exist.
with the conventional approach to design which consists in
defining and working with a succession of hierarchical layers Goal-oriented relations are useful for describing relations
of abstraction, the proof describes the construction of a layer between helper and goals strategies, while overlapping rela-
for each fundamental unit ut and us . A hidden layer in this tions describe highly connected environments, where relations
construction is the layer for the fundamental unit ub which is can enhance space exploration and learning.
used indirectly by complex rules. At the end, profitable strate- 2) Structure and Computation: Using relations between
gies will have an embodiment of the same information as W , strategies, we connect strategies into basic units of compu-
while non-profitable strategies will have a basic embodiment tation called computational digraphs (or computational con-
made only of simple rules. texts). A formal definition of a computational digraph is given
Theorem 1: (Mapping of Complex Rules to Strategies) in Definition 9 and two types of computational digraphs (local
Given a world W , a helper strategy Shi,j in W with and global) derive from it.
fp (Shi,j ) > pW , and the set Ri of all rules for Sgi , then We use SH to denote the set of all helper strategies for a
it exists a many-to-one mapping of rules from Ri to Shi,j given goal and SG to denote the set of all goal strategies.
that follows 2 steps and uses universality and embodiment
properties of Shi,j . Definition 9: (Computational Digraph) A computational
context τ in DS is a digraph which is an ordered pair
Proof: First, we determine the profit of Shi,j using its τ = (S 0 , R0 ) with
universality property, i.e., p = fp (Shi,j ). Then, for each
fundamental unit (ut and us ), we map all rules with profit • S 0 the set of all strategies (SH ∪ Sg ), called nodes.
equal to p to Shi,j . At the end, Shi,j has an embodiment of • R0 the set of ordered pairs of arcs, with R0 =
the same information that shapes W because it uses directly {(Si , Sg ) : (Si , Sg ) ∈ Rg , Si ∈ SH } or R0 =
complex rules from ut and us and indirectly simple rules from {(Sg , Si ) : (Sg , Si ) ∈ Ro , Si ∈ SH }, but not both.
ub .

871
 Definition 9 connects a strategy with a set of strategies III. M ODEL A RCHITECTURE AND AGENT D ESIGN
based on their relation, Ro and Rg and proposes a classification
In this section, we describe the hierarchical structure of the
of computation to either local or global computation. Namely,
model and offer a general solution to the agent design problem.
most of the computations in a local context are local (i.e, a
single strategy is applied at any given time) and only a few The hierarchical structure proposed by the model is made
global computations. Similarly, most of the computations in of three layers: physical reality layer, memory layer and exper-
a global context are assumed to be global and only a few of imental layer. Layers emerge from previous layers with new
them are local. information and computation capabilities. We briefly describe
each of the layers in a bottom-up fashion:
We compare local and global contexts with classical and
quantum systems. In a local context, all computations are 1) Physical reality layer. The physical reality layer
assumed to be local as in classical systems, where influences presents the physical world at a very simple level
between strategies situated in space-like separated regions are of information with extraordinary implications for
prohibited and applications of strategies depend directly on understanding the world. The first implication is that
their properties and not on the way they are applied in the we can describe the physical reality in terms of
world. A global context is similar with a quantum system, complex rules made of multiple fundamental units
where one is forced to choose a context first (i.e., a goal of information. The second implication is that we
strategy) and apply one of its rules, which immediately affects can ‘see’ the inherent connections between rules that
the application of one or more strategies (i.e., helper strategies) share fundamental units (or behavior). At this layer,
in the same context. information is represented by simple rules and com-
putation is modeled by complex rules using motion
3) Changing of Structure: We model the change of struc- and reflection operators.
ture using frames and their instances called universes. A 2) Memory layer. The memory layer consists of a set
frame represents a state in the world determined by a set of of frames, each frame describing a family of com-
computational contexts and a set of rules (Definition 10). An putational contexts. A frame uses the information
ordered set of frames describes a memory system that changes that transcends the physical layer to distribute the
its state when new rules are introduced in the world. Universes complexity of computation into different computa-
are instances of frames that evolve from a primordial universe tional contexts using goals, sub-goals and relations.
U0 , and use frames as a base knowledge to evolve strategies The computation is modeled using local and global
(Definition 11). computational contexts and it is compared with the
Definition 10: (Frames) A frame Fi is a set of computa- computation in classical and quantum systems.
tional contexts and set of rules that can be used only within 3) Experimental layer. The experimental layer makes the
Fi . It represents one evolution cycle: (a) a frame starts with a distinction between ‘reality’ and ‘experimentation’
population of strategies, (b) then, it executes mutations and using instances of frames called universes. While
selections on strategies, and (c) it reproduces strategies to frames are a knowledge base for the world, uni-
initialize a new frame Fi+1 . verses are created as a result of different types of
interactions with the world. This interpretation of
Fi = {(τi , Ri ) : τi ∈ DS , Ri ⊆ R}. (9) information allows strategies with fixed structures (as
defined in frames), to have complexity and levels of
Definition 11: (Universes) Given a frame Fi , an universe adaptation derived from evolution. It also describes
Uj in Fi is an instance that uses only a subset of the rules from
Fi and it represents a set of rule applications until a selection the formation of a ‘new reality’ that engulfs previous
occurs. The selection is followed by mutations of strategies ‘realities’ and adds new space and time dimensions.
into the next universe Uj+1 . Next, we offer a general solution to the agent design prob-
lem using the ICE agent model. The design and algorithmic
Ui = {(τi , Ri0 ) : τi ∈ DS , Ri0
⊆ Ri and
(10) steps are detailed in Figure 1. Some steps, such as 1 - a, 2 - a,
∀ri ∈ Ri0 , ai has been applied in W } b, c and d, require the existence of an interface with the agent
(human or artificial agent) that can provide useful guidance
The universe introduces new space/time dimensions: ds - is or personal creative choices, while the other steps are trivial
the space dimension that measures the set of all rules that to implement. The main operations executed by an agent are
are used in the universe, and dt - is the time dimension that movement, navigation, coordination, cooperation and learning.
measures the number of applied rules. First, the agent starts from the memory layer and describes all
When a frame transition to another frame, the structure of goals, the ordered set of frames, the set of sub-goals for each
DS is changing. The transition from a digraph DS Fi in frame goal and different types of relations. Then, the agent initializes
Fi to a digraph DS Fi+1 in the next frame Fi+1 is performed in an active memory with a current population of solutions that
two steps: (1) an evolutionary process reproduces all strategies are subjected to different types of operations. Operations are
in DS Fi and connects any two children strategies with the grouped in four modules - selection, variation, learning and
same type of arcs as their parents, producing DS 0Fi , and (2) a control modules.
new digraph DS 00Fi+1 is created as the union of all nodes and The selection module describes the selection for reproduc-
arcs from DS 0Fi and a new digraph DS Fi+1 that defines new tion and mutation: selection for reproduction (3 − b − i) allows
goals and sub-goals for Fi+1 . all individuals to reproduce in order to create a new, improved

872
 1) Construct the physical layer.
reality resulted from evolution and learning, and selection for
mutation (3−b−ii) is the result of coordination with opponent a) Given a world W , describe the rules of the world.
b) Associate rules to fundamental units using param-
individuals having a similar embodiment. By coordination we
eters pW and cW .
imply the use of a common fitness function for all agents
2) Construct the memory layer.
to compare individuals with similar embodiment. In addition,
agents can use individual ranking functions based on profit / a) Describe all goals of W and associate appropriate
rules to each goal.
complexity / frequency of actions in a given universe.
b) Organize goals into families of goals or frames.
The variation module defines the mutation (3 − c − i) and c) Order frames based on their importance / rele-
reproduction (3 − c − ii) operators using different types of vance.
d) For each goal, choose a set of sub-goals up to a
adaptations. Cooperating agents can mutate individuals with given profit bound and construct each sub-goal.
random opponent rules of best opponent individuals. Repro-
3) General structure of an evolutionary algorithm.
duction of individuals creates new individuals with different
adaptations to a new frame. For example, the reproduction of a) Use an active memory that contains the current
population of individuals.
a single individual can produce two types of new individuals: b) Use a selection module to describe the selection
clones with minimum adaptation and well adapted individuals for reproduction and mutation in a population of
having all the rules of the parent and, in addition, the entire individuals: (i) selection for reproduction allows
set of basic rules of the new frame. all individuals to reproduce and (ii) selection for
mutation is the result of coordination with oppo-
The learning module proposes the creation of a new reality nent individuals with similar embodiment.
resulted from the cooperation and coordination between differ- c) Use a variation module to describe the creation of
ent agents (3 − d − i), from local interactions with individuals new individuals by means of mutation and repro-
from the same population (3 − d − ii) and the incorporation duction: (i) use cooperation to mutate individuals
of best experiences (3 − d − iii). Through cooperation and with random opponent rules of best opponent
coordination agents learn best opponent individuals. In addi- individuals or take random rules from the memory
tion, local interactions are important for increasing the speed layer and (ii) use different types of adaptation to
of learning and the exploration of emergent behavior. reproduce individuals in new frames.
d) Use a learning module to describe learning situ-
The control module uses the morphology of individuals ations: (i) use cooperation and coordination with
to select the most profitable individuals (3 − e − i) and, for opponent individuals with similar embodiment, (ii)
a given individual, the most profitable or the most complex use local interactions with individuals from the
same population to increase the speed of learning
actions (3 − e − ii). and explore emergent behaviors, and (iii) use its
Concluding, the successful design of autonomous agents own experience to learn from the past universes.
e) Use a control module to describe search opera-
relies on the effective embodiment and the emergent behavior.
tions: (i) use the morphology of individuals to
Self-organization of individuals distributes rules among strate- select the most profitable individuals or (ii) the
gies with different profits. Collective behavior emerges through most profitable or the most complex actions.
the process of self-organization, from the local interactions
between individuals that share a relation. Fig. 1. Agent design using the ICE agent model.

IV. C ONCLUSION AND F UTURE WORK R EFERENCES

[1] A. E. Eiben and J. E. Smith, Introduction to Evolutionary Computing.
This work contribute to the modeling of agents by taking Berlin, Germany: Springer-Verlag, 2003.
inspiration from both physics and biology. It describes the [2] J. M. Epstein, Generative Social Science: Studies in Agent-Based Com-
world as operating at a very simple level of information using putational Modeling. Princeton, NJ: Princeton University Press, 2006.
basic rules, from which the complexity of the world emerges in [3] D. Floreano and C. Mattiussi, Bio-Inspired Artificial Intelligence: Theo-
the form of more complex rules. The computation is described ries, Methods and Technologies. Cambridge, MA: MIT Press, 2008.
using structures (i.e, computational contexts and memory [4] V. Vedral, Decoding Reality: The Universe as Quantum Information.
of successive configurations of computational contexts) and University of Oxford, United Kingdom: Oxford University Press, 2010.
objects (i.e, strategies). The work concludes with a general [5] H. Zenil, A Computable Universe: Understanding and Exploring Nature
solution to the agent design problem. as Computation. World Scientific Publishing Company, Singapore:
World Scientific, 2013.
The current work can be extended in several important [6] T. Ingold, “An anthropologist looks at biology,” in Man, 1990, vol. 25,
directions: 1) A promising direction of the current research no. 2, pp. 208−229.
is to extend the model with more operators that capture [7] R. Pfeifer, M. Lungarella and F. Iida, “Self-organization, embodiment,
and biologically inspired robotics,” in Science, 2007, vol. 318, no. 5853,
more complex ideas in physics (e.g., gravitational fields or pp. 1088−1093.
gravitational potential energy); 2) Another direction is to
[8] S. Wolfram, A New Kind of Science. Champaign, IL: Wolfram Media
implement the model for different learning scenarios (e.g., Inc., 2002.
learning representative instance of text or learning how to play [9] H. Zenil, “Introducing the Computable Universe,” in CoRR, jun, 2012.
games); 3) Finally, the ultimate goal is to capture behavioral [10] H. Zenil, “Information Theory and Computational Thermodynamics:
characteristics of human subjects and to explore and under- Lessons for Biology from Physics,” in Information, 2012, vol. 3(4), pp.
stand the dynamics underpinning brain function. 739−750.

873