Conference PaperPDF Available

Coevolutionary Learning of Neuromodulated Controllers for Multi-Stage and Gamified Tasks



Neural networks have been widely used in agent learning architectures; however, learning multiple context-dependent tasks simultaneously or sequentially is problematic when using them. Behavioural plasticity enables humans and animals alike to respond to changes in context and environmental stimuli, without degrading learnt knowledge; this can be achieved by regulating behaviour with neuromodulation – a biological process found in the brain. We demonstrate that modulating activity-propagating signals when evolving neural networks enables agents to learn context-dependent and multi-stage tasks more easily. Further, we show that this benefit is preserved when agents occupy an environment shared with other neuromodulated agents. Additionally we show that neuromodulation helps agents that have evolved alone to adapt to changes in environmental stimuli when they continue to evolve in a shared environment.
Coevolutionary Learning
of Neuromodulated Controllers
for Multi-Stage and Gamified Tasks
Chloe M. Barnes, Anik´
o Ek´
art, Kai Olav Ellefsen, Kyrre Glette†‡, Peter R. Lewisand Jim Tørresen†‡
Department of Computer Science, Aston University, Birmingham, UK
{barnecm1, a.ekart, p.lewis}
Department of Informatics, University of Oslo, Oslo, Norway
{kaiolae, kyrrehg, jimtoer}
RITMO, University of Oslo, Oslo, Norway
Abstract—Neural networks have been widely used in agent
learning architectures; however, learning multiple context-
dependent tasks simultaneously or sequentially is problematic
when using them. Behavioural plasticity enables humans and
animals alike to respond to changes in context and environmental
stimuli, without degrading learnt knowledge; this can be achieved
by regulating behaviour with neuromodulation – a biological
process found in the brain. We demonstrate that modulating
activity-propagating signals when evolving neural networks en-
ables agents to learn context-dependent and multi-stage tasks
more easily. Further, we show that this benefit is preserved when
agents occupy an environment shared with other neuromodulated
agents. Additionally we show that neuromodulation helps agents
that have evolved alone to adapt to changes in environmental
stimuli when they continue to evolve in a shared environment.
Natural and artificial environments are often complex,
unpredictable and dynamic, making learning and surviving
a challenge for both animals and artificial agents alike [1],
[2]. In order to survive in these challenging conditions, many
organisms such as nematodes [3], fish [4] and the African
striped mouse Rhabdomys pumilio [5] show phenotypic, ac-
tivational and behavioural plasticity; this ability to express
different behaviours – and reverse them depending on varying
environmental stimuli – allows rapid adaptation to novel
situations [5], [6].
As with animals in the real world, controllers tasked with
learning in dynamic and unpredictable artificial environments
– artificial neural networks (ANNs) in particular – face similar
challenges. Neuroevolution is the process of evolving ANNs
with an evolutionary algorithm in accordance to a fitness func-
tion [7]; a population of individuals is evolved with mutations
and/or crossover over many generations [8]. Many applications
of neuroevolution focus on evolving weights of ANNs [8]–
[11], however more complex approaches that evolve both the
weights and topologies of ANNs exist [12], [13]. This process
of evolving ANNs by adjusting connection weights over time
to encode new information can result in a degradation of per-
formance and catastrophic forgetting when learning new tasks
or experiencing novel environmental contexts [8], [14]–[16];
learnt knowledge must be changed – and is often lost – in order
to learn new things and express new behaviours [8]. Learning
complex, sequential or multi-stage tasks is also made difficult
as complete information about the environment – including
the available actions, their cues and their consequences – is
not usually accessible [1], [17]; this is also evident when
environments are shared, as the actions of individuals change
the context of the environment for others [11].
In nature, the immediate and reversible behavioural changes
as a result of behavioural plasticity that facilitate adaptations
to novel contexts can be achieved with neuromodulation
a biological process whereby chemical signals are gated in
the brain depending on environmental stimuli and situations
[18]. Consequently, neuromodulation has thus been used to aid
neural controllers with learning new or sequential tasks, and
learning in dynamic environments [8], [19], [20].
We explore how activity-gating neuromodulation may help
neural controllers to overcome the challenges associated with
learning multi-stage and gamified tasks, without a priori
knowledge of the task or environment. ANNs are just one
example of an agent controller in which behaviour can be
learnt; we use ANNs in this context in line with previous
River Crossing testbeds [9]–[11], to explore how ANNs make
decisions in social environments. Here, a multi-stage task is
defined as one that an agent must learn, and pass, through
multiple states, and perform different behaviours in different
contexts in order to achieve their goal; this definition is
inspired by [17]. Further, we investigate how this regulation
of behaviour may help these agents to learn in multi-agent
environments, without the capacity to learn of the existence of
others; the act of introducing other agents to the environment
changes the context of the task, which becomes an implicit
social dilemma. Neuromodulation has been used to explore
social dynamics in multi-agent systems [2], [21], however
our work extends the notion of [11], where cooperation
and exploitation may be emergent behaviours, but cannot be
intentional. We hypothesise that reversible and immediate be-
havioural plasticity as a result of neuromodulation will enable
agents to better learn the state-space of their environment
and therefore the task at hand. We demonstrate this using
the River Crossing Dilemma (RCD) testbed introduced by
[11], as well as a new adaptation of the environment called
the Protected River Crossing Dilemma (PRCD), which we
introduce to explore and contrast how agents learn to solve
single-stage tasks under these same conditions.
A. Behavioural Plasticity and Neuromodulation
One way to design adaptive systems is to look at the theory
of behavioural plasticity. Behavioural plasticity can be seen
as the ability to change or adapt behaviour based on changes
in environmental stimuli [22]; this is important for navigating
uncertain, novel or dynamic environments and can be classed
into two different types: developmental and activational [6].
Developmental behavioural plasticity can be seen as learning
from experience and external stimuli. Activational behavioural
plasticity on the other hand enables immediate behavioural
changes; individuals can respond to new or dynamic environ-
ments during their lifetime by changing their phenotype. Ac-
tivational plasticity is also termed ‘innate’ [22] or ‘contextual’
[23] plasticity.
Neuromodulation is a biological process found in ani-
mal brains [24], whereby chemical signals modify, gate or
regulate synaptic plasticity based on the modulatory signal
combined with the pre- and post-synaptic activities, and en-
vironmental stimuli [8], [18], [25]. In neuroscience, synaptic
plasticity is the modification of synapses between neurons
through strengthening or weakening them [26]. In ANNs,
synaptic plasticity is achieved by modulating neural network
weights; short-term modifications result in immediate pheno-
typic changes, and long-term changes result in learning and
adaptation based on experience. Developmental plasticity is
achieved by regulating learning in the long-term, where mod-
ulatory signals alter synaptic strengths; activational plasticity
is achieved by regulating behaviour or synaptic activity in the
short-term with the modulatory signal, without affecting learn-
ing and without long-lasting changes to synaptic strengths.
B. Achieving Developmental Plasticity with Neuromodulation
Similarly to ANNs being inspired by the connectionist
architectures found in brains, neuromoduation has been widely
applied to artificial models to regulate synaptic plasticity and
the learning rate of neural connections. Neural networks have
been evolved with modulatory neurons to regulate learning and
mitigate the catastrophic forgetting associated with performing
tasks in uncertain environments [25]; this method has been
found to improve learning in T-maze problems. Other studies
have found that promoting the evolution of modular neural
networks by introducing a cost for neural connections can
mitigate catastrophic forgetting and improve learning; here,
learning is regulated with neuromodulation [8]. Neuromodula-
tion has also been used to develop conflict learning in neural
networks [27], and associative learning in real robots [28];
these two approaches employ neuromodulation, but do not use
neuroevolution as a learning mechanism.
The approaches outlined in this section modulate learning
and therefore developmental plasticity by regulating the local
learning rate of neurons in the network; they do not however
demonstrate how behaviour can be regulated in a short-term,
reversible way without affecting learning, in order to facilitate
immediate behavioural changes to changing environmental
stimuli. Further, these approaches only use neuromodulation
in neural networks or robots that exist in isolation; we however
explore how immediate behavioural plasticity can be achieved
with neuromodulation in agents without regulating learning,
in single- and also multi-agent environments.
C. Achieving Activational Plasticity with Neuromodulation
Neurobiological mechanisms have been explored using a
computational framework based on neuromodulatory systems
such as the dopaminergenic and serotonergic systems, by
regulating synaptic activity [29]. Whilst this is proposed to
aid autonomous agents in exploratory and exploitative decision
making, activational plasticity is not applied as a tool to
improve neuroevolution, but rather to model and explore
biological systems computationally. The effects of modulating
neuroreceptors and synaptic plasticity have been studied with
spiking neural networks to model EEG data [30]; an aim
of this work is to produce a tool to explore and diagnose
neurological disorders such as dementia – and not to use
neuromodulation as a tool to aid artificial agents in achieving
goals. Supervised learning methods and ‘context-dependent
plasticity’ – or ‘activational plasticity’ [6] – have been shown
to be beneficial for maintaining high accuracy for large num-
bers of sequential classification tasks, based on the MNIST
and ImageNet datasets [31]; this was achieved by gating
activations randomly in the network for each task. In other
work, ‘context-dependent selective activation’ is achieved by
learning parameters of a separate neuromodulatory network,
which in turn gates activity for a prediction network [32];
this two-layered neural network approach is used for learning
sequential tasks. This approach indirectly modulates learning,
as the amount of activity in the predictive network after
modulation is reflected in the back-propagation process.
Whilst it is common for learning and activity to be regulated
by a separate group of modulatory neurons or an entire
network [19], [20], [32], a distinguishing characteristic of our
work is that we explore the impact that regulating activity-
propagating signals within a single neural network has on an
agent’s ability to learn multi-stage tasks. By not explicitly
regulating learning, we regulate behaviour to provoke im-
mediate phenotypic changes based on environmental stimuli.
Additionally, we use neuroevolution to evolve which neurons
in the neural network are modulatory, resulting in a more
structured way of operationalising neuromodulation than [31]
for example, where neuronal activity is gated randomly.
D. Learning Multi-Stage Tasks in Multi-Agent Environments
Both humans and animals find learning in environments that
change state or context without explicit cues challenging - this
however is a characteristic of most realistic environments [1];
these changes need to be detected in order to adapt behaviour
accordingly, as it is rare for this information to be explicitly
available. This has also been identified as being a difficulty of
learning multi-stage tasks, as the full state-space of tasks is
not usually available when learning [17]; changes in state or
stimuli also change the context in which behaviours are learnt.
These challenges are also present when neural networks
learn to achieve new or many tasks, or navigate dynamic or
uncertain environments; encoded knowledge must be adapted
in order to learn new things [8].
Regulating synaptic plasticity with neuromodulation has
been shown to facilitate adaptation and learning when there
are changes in environmental stimuli or the task at hand, thus
helping agents to overcome these issues [8], [16], [20], [25].
Whilst neuromodulation has also been used in multi-agent
contexts, this is typically to explore the effect on cooperative
or competitive strategies in social dilemmas [21] or in
competitive environments [2], where agents are explicitly
aware of others and thus employ strategies intentionally.
Agents acting in novel environments may not have full or
even partial information about others in the environment, and
thus cannot cooperate or compete intentionally. In previous
work, we show that learning in multi-agent environments
without knowledge of the existence of others is problematic,
as environmental stimuli change unpredictably as a result of
the actions of others in the environment [11]. Social action is
shown to improve learning in agents situated in multi-agent
environments [11], however agents do not exhibit behavioural
plasticity or use neuromodulation; furthermore, the study is
limited to exploring multi-stage tasks.
In this paper, we aim to explore the challenges presented
to neural controllers when they experience unpredictable en-
vironmental stimuli, and observe the effect that behavioural
plasticity arising from the regulation of activity-propagating
signals has on learning. As seen in the natural world [3],
we would expect agents capable of behavioural plasticity to
adapt better to changing, dynamic and uncertain environments,
than those that are not. We investigate this by evolving agents
to learn single- and multi-stage tasks, in both single- and
multi-agent environments; this covers different combinations
of environmental changes and variations. Specifically, we
use the term ‘multi-stage task’ in a similar context to [17],
where agents must learn multiple stages of a task in order to
achieve a goal. Further, we explore the effect that changing
the context in which an agent exists has on evolution, by
evolving agents in an environment alone for a period of time
and then evolving them for additional time within a multi-
agent environment. We hypothesise that activational plasticity
will help agents to achieve their tasks in these environments,
by facilitating immediate behavioural changes in response to
different environmental contexts or conditions.
A. The River Crossing Dilemma Testbed
The River Crossing Dilemma (RCD) testbed was introduced
by Barnes et al. [11], to explore how agents evolve to achieve
individual goals in shared worlds; this extends the original
River Crossing Task proposed by Robinson et al. [9]. Agents
have no prior knowledge of the task or environment, and must
learn what their goal is and how to achieve it without this
information. The RCD is a 19 ×19 grid-world, with a two-
cell deep river of Water in the centre. There are four Stones on
each river bank, and all empty cells are Grass. An agent’s goal
is to collect one of its allocated Resources from either side of
the river, rewarding the agent with a highly positive fitness.
Conversely, stepping into the river causes the agent to drown,
giving it a highly negative fitness. As a result, the task is multi-
stage [17], as agents must evolve to perform sub-tasks and the
appropriate behaviours that correspond to different states and
environmental stimuli – they must build a bridge to cross the
river to achieve their goal. As the river is two cells deep, two
Stones must be placed in the same Water cell to successfully
build a bridge; agents must step onto a cell with a Stone to pick
it up, then stand adjacent to the river to partially or fully build
a bridge. The RCD testbed is a bespoke Java implementation,
and is presented in Figure 1. Time is measured in ‘timesteps’,
where an agent can move a distance of one cell per timestep.
For experiments with two agents, the agent that starts in the
top left of the environment moves first, followed by the agent
that starts in the bottom right.
Fig. 1. The River Crossing Dilemma testbed, proposed by Barnes et al. [11].
The grey agent (top left) is allocated the two Resources in grey, and the black
agent (bottom right) is allocated the two Resources in black; agents cannot
interact with Resources not allocated to them. Both agents can interact with
all other objects. For single-agent environments, the black agent is removed.
B. The Protected River Crossing Dilemma
We introduce the Protected River Crossing Dilemma
(PRCD) – an adaptation of the RCD [11] specifically used
to explore how agents evolve to solve single-stage tasks; like
with the RCD, the PRCD is a bespoke Java implementation.
The environment is constructed as seen in Figure 1, but the
river acts as an impassable – and most importantly, a non-
lethal – obstacle, meaning agents cannot fall into it mistakenly.
This simple change means that agents do not need to learn the
different states in which they can interact with the river: that it
is not safe unless the agent is carrying a Stone. As the PRCD
river is impassable, agents must still perform sub-tasks such as
bridge-building to succeed; removing the river entirely would
remove the multi-stage task but also make the task trivial.
The single-stage task therefore reduces the variability in the
task and environment, making it less complex; as plasticity is
said to increase with environmental variability [33], we would
expect the effect of neuromodulation to be less apparent than in
the multi-stage RCD task. Further, we would expect that the
benefit of neuromodulation is less evident still when agents
evolve to solve the single-stage task alone compared to when
they evolve together for this same reason.
C. Gamification of the RCD and PRCD
The RCD and PRCD are gamified, such that agents incur an
increasing cost for each Stone placed in the river; a bridge is
successfully built with two Stones. Therefore, in multi-agent
environments, agents face a social dilemma and may either
complete their task individually and be subjected to the full
cost of bridge-building, cooperate to share the cost, or exploit
other agents to avoid a cost at all. This creates a Snowdrift
Game [34] (also known as the Chicken Game [21], [35]–[37]
or the Hawk-Dove Game [21], [38]), resulting in less incentive
to cooperate due to the cost of bridge-building, but failure for
defection if the agent isn’t able to achieve its goal. The fitness,
or payoff, for agent piis calculated with Equation 1:
21 + sif(1)
where ris the number of Resources collected by pi,N= 2
and is the number of Resources allocated to each agent,
C= 0.1and is the cost of placing a Stone in the river, si
is the number of Stones placed in the river by agent pi, and
f= 1 if agent pifalls in the river, or 0otherwise. An agent’s
fitness therefore records its own behaviour.
Commonly observed fitnesses are presented in a payoff
matrix in Table I, using Equation 1. The maximum fitness
an agent can achieve alone is 0.7, which increases to 0.9if
the cost of bridge-building is shared, or 1.0if an agent exploits
another in a shared environment; anything below 0.7indicates
the goal is not achieved.
D. Agent Design
Agents in both the RCD and the PRCD use a two-layered
neural network architecture, adapted from [11] and inspired by
[9]. The deliberative layer generates high-level sub-goals based
on the current inputs, corresponding to the agent’s current
state. This network is therefore responsible for the decision-
making processes of agents; depending on the inputs and the
weights of the network, the outputs indicate what the agent
decides to do next in terms of sub-goals (whether it is attracted
Agent 1
Agent 2 S2= 0 S2= 1 S2= 2
S1= 0 0.0
S1= 1 -0.1
S1= 2 0.7
to, neutral towards or repulsed from certain objects in the
environment). The weights of the network (as well as the type
of each neuron in the network) therefore represent the genes
of the agent, and therefore what behaviours it will exhibit
depending on what inputs. The inputs are 1or 0depending
on whether the agent is on Grass, a Resource, Water or a
Stone, if it is currently carrying a Stone, and if a bridge has
been built partially in the environment (i.e. one Stone in the
river out of two). This fully-connected feed-forward network
has six input neurons, three hidden layers with eight, six and
four neurons respectively, and an output layer of three neurons
(Figure 2). Resources, Stones and Water will be attractive if
the output is 1, avoided if 1, or neutral if 0.
Fig. 2. The Deliberative Layer is a fully-connected neural network with
three hidden layers, that generates high-level sub-goals. Inputs are 1or 0,
corresponding to the agent’s current state: Grass, Resource, Water, Stone,
Carrying Status, if a Bridge partially exists. Outputs are 1for attraction, 0
for neutral or 1for avoidance for each sub-goal: Resource, Stone, Water.
The reactive layer is a neural network with the same dimen-
sions as the environment – in this case, 19 ×19 – where each
neuron is connected to the surrounding eight neurons. This
reactive neural network uses the shunting equation (Equation
2, [9], [11], [39], [40]) to create dynamic activity landscapes
at each timestep based on the current sub-goals; agents can
therefore hill-climb towards the goals generated in the previous
layer by moving to the cell in its Moore neighbourhood (the
surrounding eight cells) with the highest activity. Agents must
make one move per timestep – they cannot remain stationary.
Additionally, an agent will pick up a Stone automatically if
it moves onto a cell with a Stone; an agent will also put
a Stone in the river automatically if the cell to its left or
right is Water – and if it is carrying a Stone. Equation 2
calculates the activity of each neuron based on its own and
the surrounding activations: Ais the passive decay rate; xiis
the current neuron; wij is the weight between neurons xiand
xj, where xjis one of the surrounding cells in xi’s Moore
neighbourhood (indicated by k= 8); [xj]+is calculated by
max(0, xj), meaning that negative activity cannot propagate
through the network. Iis the Iota value of the neuron, which
depends on the sub-goals from the deliberative layer (for a
value of: 1,I= 15;1,I=15; and I= 0 otherwise); this
creates hills and valleys in the activity landscape, as inspired
by the original RCT testbed [9].
dt =Axi+Ii+
wij [xj]+(2)
E. Operationalising Activity-Gating Neuromodulation
Neuromodulated agents regulate their behaviour by gating
activation within the neural network in the deliberative layer
(Figure 2); this distinguishes our approach from others, which
either use a separate modulatory network/neurons, or regulate
learning as well as, or instead of, behaviour [19], [20], [32].
Figure 3 shows an example of this activity-gating modu-
lation. Neurons in the deliberative layer may evolve to be
non-modulatory or modulatory; if the incoming signal to
a modulatory neuron is negative, it will fire and regulate
behaviour by outputting a signal of 0 on each of its outgoing
connections. This means that weights on the connections
will effectively be ‘turned-off’, or gated, as the signal is
blocked locally. Immediate, and more importantly reversible
behavioural changes can therefore be achieved depending on
the stimuli experienced. This gating or modulation of activity-
propagating signals results in behavioural plasticity; an agent’s
genotype, represented by the evolved weights of the neural
network and the types of the neurons in the deliberative layer,
Non-Modulatory Neuron
Modulated Connection
Modulatory Neuron
Fig. 3. Activity-Gating Neuromodulation: Modulatory neurons propagate
activity the same as non-modulatory neurons when the incoming signal is
0; here, if the incoming activity signal to x2is positive, the outgoing
activity signals of x2propagate as usual and are passed on to the next layer of
neurons (in this case, y1and y2). If, however, the incoming activity signal to
x5is negative, the modulatory neuron fires and the outgoing activity is gated;
specifically, this means that the neuron x5will output signals of 0along each
of its outgoing connections (in this case to y3and y4), so the outgoing signal
is effectively gated or ‘turned off’ when the signal is multiplied by the weight
of the connection. This means agents can exhibit behavioural plasticity, as the
weights of the neural network are not changed, but temporarily suppressed;
this leads to the network producing different outputs and therefore different
behaviours, without modifying the network weights in a permanent way. It
is important to note that modulatory neurons only affect their own outgoing
connections, so the connections from x4and x6to y3and y4are not affected
when x5fires.
is therefore able to express multiple phenotypes depending
on state and environmental stimuli – without changing, or
potentially destroying, the knowledge encoded in the weights
of the network. In other words, one deliberative neural network
with modulatory neurons can exhibit temporary and reversible
changes in behaviour depending on the stimuli and inputs; this
is because modulatory neurons that are ‘switched off’ do not
propagate any activity signals to the next layer of neurons, thus
changing the output of the network and the resulting behaviour
of the agent.
F. Evolutionary Algorithm
All experiments are conducted using the RCD testbed with
the following common parameters, inspired by [11]. For each
experiment, a population of 25 randomly initialised agents
is evolved using a Steady State Genetic Algorithm. Agents
acquire knowledge, and therefore ‘learn’, through evolution
– there is no within-lifetime learning. At each generation,
three agents are randomly selected from the population and
are evaluated in a tournament; each agent has 500 timesteps
to navigate the environment and achieve their goal. The
evaluation stops if all agents reach the maximum amount
of timesteps, achieve the goal, or die. The agent with the
worst fitness in each tournament is replaced with an offspring
generated from the best two. For each chromosome (layer
of weights in the deliberative layer), this offspring has a
probability of Pone = 0.95 to inherit the chromosome from a
random parent, otherwise single-point crossover is used. Each
connection weight win the offspring’s deliberative layer is
then mutated by a random value from a Gaussian distribution
with µ=wand σ= 0.01.
For neuromodulatory agents, the neurons in the hidden lay-
ers of the deliberative neural network are evolved in addition to
the weights (input and output neurons cannot be modulatory);
neurons may evolve to be standard non-modulatory neurons, or
activity-gating modulatory neurons. The type of each neuron
(modulatory or non-modulatory) is therefore not specified in
advance, but evolved with neuroevolution like the weights of
the network. The deliberative neural network of each agent
is initialised with only non-modulatory neurons at the start
of evolution. At each generation, the new offspring inherits
the neuronal structure from a randomly chosen parent, where
the parents are the two agents with the best fitnesses in
the tournament as described above; there is a probability
of Pmut = 0.15 that one randomly chosen neuron out of
the three hidden layers in the deliberative network (Figure
2) will be mutated, from non-modulatory to modulatory or
vice versa. This mutation rate is adapted from the mutation
operators and probabilities used in [8]. Modulatory neurons
regulate activity as outlined in Section III-E. Agents that do not
use neuromodulation have a static network of non-modulatory
neurons that do not evolve.
The experiments in this study aim to investigate the effect
that behavioural plasticity through activity-gating neuromodu-
lation has on agent evolution when the environment is prone
to change; we use a series of experiments, outlined below,
to explore the extent to which the ability to rapidly and
reversibly change phenotypic behaviour helps agents to solve
tasks in varying environmental conditions. All experiments are
repeated 100 times using the same 100 seeds, both with and
without neuromodulation. The first four sets of experiments
evolve agents for 500,000 generations from a randomly-
initialised state; the final set of experiments evolves agents
for 1,000,000 generations in total by first evolving agents in
an environment alone, and then continuing to evolve together.
The first set of experiments explore how agents evolve
to solve a single-stage task in the Protected River Crossing
Dilemma (PRCD), when they exist alone in the environment.
This environment has the least inherent variability, which will
provide a baseline to compare the effects of neuromodulation
in the later experiments.
The second set of experiments introduces another agent
into the single-stage task PRCD environment, which creates a
social dilemma as agents may evolve to cooperate or exploit
the other unintentionally. As agents cannot perceive or reason
about the actions or existence of other agents, their environ-
ment appears unpredictable and therefore harder to evolve in.
These experiments evolve two separate, randomly-initialised
populations of agents that start on opposite corners of a shared
PRCD world. In these experiments and the others that involve
multi-agent environments, only the agent that begins in the
top-left corner is assessed, so all results are comparable.
The third set of experiments investigates how agents that
exist alone evolve to solve a multi-stage task, by instead using
the RCD environment. This also adds an element of variability
and uncertainty compared to the first set of experiments.
The fourth set of experiments use the RCD environment to
explore how agents that share an environment together evolve
to solve multi-stage tasks. Of these four experiments, this
environment is the most variable, due to the imperceptible ac-
tions of the other agent within the environment, as well as the
challenge of evolving to solve the multi-stage task. We expect
to observe the most pronounced benefit of neuromodulation
and behavioural activity in these experiments, as behavioural
changes are increasingly useful as environmental conditions
change [3].
The final set of experiments adds further unpredictability
to the task; here, we explore how agents that have evolved
alone for an initial period of 500,000 generations are able to
achieve their goals and adapt when they continue to evolve in
a shared environment with another agent for a further 500,000
generations. Agents therefore evolve for a total of 1,000,000
generations, with the initial period of evolving alone being
identical to the first set of experiments. Here, we explore the
extent to which activity-gating neuromodulation affects how
agents evolve to solve multi-stage tasks, when the environment
explicitly changes context from a single- to a multi-stage
environment – this is the most challenging of the five sets
of experiments due to the increase in environmental changes.
A. Learning Single-Stage Tasks When Alone
We start by investigating how agents are able to evolve to
solve the simplest task in the least variable environment in the
study – the single-stage task in the Protected River Crossing
Dilemma (PRCD) – and the role that neuromodulation plays.
Figure 4(a) shows the mean best-in-population fitness over
time when agents evolve alone in the PRCD, both with
and without neuromodulation. Here, neuromodulation appears
to be increasingly beneficial over the course of evolution.
Once the effect of neuromodulation is sustained, there is
a clear benefit to behavioural plasticity in this environment
once agents have finished evolving; 85% of agents were able
to solve the single-stage task and achieve their goal with
neuromodulation, compared to only 40% of agents that did
not use neuromodulation (Table II).
B. Learning Single-Stage Tasks When Together
By introducing two agents into the single-task PRCD envi-
ronment, the variability of the environment increases, and the
task becomes gamified. Agents may evolve to achieve their
goal alone, cooperate unintentionally, or exploit the actions of
the other agent; agents therefore have the potential to achieve
a higher fitness, at the risk of relying on the actions of another
to achieve their goal. Agents are unable to perceive others or
their actions, so the environment becomes unpredictable when
it is shared with another agent.
Figure 4(b) shows the mean best-in-population fitness of
agents evolving together in a shared PRCD environment.
Similarly to when agents evolve to solve a single-stage task
alone (Figure 4(a)), the effect of neuromodulation is slow to
Experiment Task Fitness (% of Agents)
(S/M) 0.7 0.9 1.0 <0.7 0.7
Alone S 40 0 0 60 40
Alone with NM S 85 0 0 15 85
Alone M 37 0 0 63 37
Alone with NM M 77 0 0 23 77
Together S 29 5 27 39 61
Together with NM S 49 2 46 3 97
Together M 27 5 36 32 68
Together with NM M 44 0 50 6 94
CE M 40 1 32 27 73
CE with NM M 47 2 50 1 99
0 1 2 3 4 5
Generation (100,000s)
With NM
(a) Single-Stage Task – Alone
0 1 2 3 4 5
Generation (100,000s)
With NM
(b) Single-Stage Task – Together, Gamified
0 1 2 3 4 5
Generation (100,000s)
With NM
(c) Multi-Stage Task – Alone
0 1 2 3 4 5
Generation (100,000s)
With NM
(d) Multi-Stage Task – Together, Gamified
Fig. 4. The mean best-in-population fitnesses of agents evolving to solve (a) a single-stage task alone, (b) a single-stage task together, (c) a multi-stage task
alone and (d) a multi-stage task together, for 500,000 generations, with and without neuromodulation (NM). Single-stage tasks take place in the PRCD, and
multi-stage tasks take place in the RCD. A fitness of: 0.7 indicates the goal is achieved individually; 0.9 indicates the cost of bridge-building is shared; 1.0
indicates an agent exploits another’s act of building a bridge; 0.7 or above indicates the goal is achieved; below 0.7 indicates the task is failed (Equation 1).
manifest; however, as agents can access a higher fitness than
they can achieve alone, the effect of neuromodulation is more
prominent than when agents are alone. In Figure 4(b), agents
evolve to achieve a higher fitness more often, and by the end of
evolution, 97% of neuromodulatory agents achieve their goal
compared to 61% of non-modulatory agents.
C. Learning Multi-Stage Tasks When Alone
The multi-stage task present in the RCD creates a more
variable environment than that seen in the single-stage task
PRCD environment; agents must evolve to match correct
behaviours with different environmental stimuli under different
conditions, which is a more challenging – and more perilous
– task when the possibility of falling in the river exists.
When agents evolve alone in the RCD environment, they
can only achieve their goal once they have built a bridge
on their own. As the environment is gamified, the maximum
fitness an agent can achieve is therefore 0.7, after the bridge-
building cost is deducted from the fitness (Equation 1).
The mean best-in-population fitness increases over time
as more agents evolve successful solutions; after 500,000
generations, 37% of agents evolved the necessary behaviours
to achieve their goal without neuromodulation, compared to
77% that achieved their goal with neuromodulation (Table
II). Figure 4(c) shows that the mean best-in-population fitness
increases faster when agents use neuromodulation, indicating
that agents are more likely to evolve successful solutions,
and that they are able to do this in fewer generations than
agents that do not use neuromodulation. The increase in task
complexity and environmental variability compared to the
single-stage task PRCD environment indicates that behavioural
plasticity is more beneficial when there is greater variability
or uncertainty in the environment or the task at hand.
D. Learning Multi-Stage Tasks When Together
The fitness function presented in Equation 1 evaluates
each agent individually. When agents share an environment,
they can still achieve their goal alone by building a bridge
completely by themselves and enduring the associated cost, but
they can also exploit the other to avoid the cost, or cooperate
and share the cost of bridge-building. The maximum accessible
fitness to each agent therefore increases to 1.0instead of
0.7, as agents may achieve their goal without building a
bridge. In each case, agents have no capacity to perceive
the existence or actions of the other, so cannot cooperate or
exploit intentionally; instead, these agents perceive changes in
environmental stimuli, and attempt to adapt their behaviour
0 1 2 3 4 5 6 7 8 9 10
Generation (100,000s)
With NM
Fig. 5. The mean best-in-population fitness of agents that evolve alone for 500,000 generations, then continue to evolve together with a random partner for
a further 500,000 generations, with and without neuromodulation (NM). A fitness of 0.7 or above indicates the goal is achieved (Equation 1).
The multi-stage task presented in the RCD adds yet another
layer of complexity onto the task and the environment; a multi-
agent environment introduces an element of unpredictability
as agents cannot perceive others, and a multi-stage task
means that the agent must discover multiple states and the
corresponding consequences in the environment in order to
achieve its task.
Figure 4(d) shows that neuromodulated agents that regulate
their behaviour evolve to achieve their goal more often, and in
fewer generations, than non-modulated agents. After 500,000
generations, 94% of neuromodulated agents achieve their goal,
compared to only 68% of non-modulated agents (Table II).
This shows that agents receive a benefit from expressing
behavioural plasticity in response to changes in environmental
stimuli caused by the actions of others.
E. Learning a Multi-Stage Task with Continued Evolution
When agents evolve alone in the RCD, the maximum fitness
they can achieve is 0.7after the total cost of building a
bridge is deducted. When agents evolve together, this threshold
increases to 1.0as the possibility to exploit the bridge-
building of other agents arises. In the following experiments,
agents undergo an initial period of evolution in the multi-
stage RCD environment alone for 500,000 generations. After
this initial period of evolution, agents are then paired with
another agent who has also evolved alone, and both continue to
evolve together in a shared, multi-stage task RCD environment
for a further 500,000 generations. By changing the agents’
environment from an individual to a shared environment, the
predictability decreases not only because the environment is
now shared – but because the context in which the agents have
evolved in is completely changed. Agents must adapt their
behaviour to cope with a change in environmental stimuli,
and the unanticipated actions of others in the environment.
Figure 5 shows the full evolution of the agents that evolve
for a period alone (generations 1-500,000), and then continue
to evolve in a shared environment with another agent (gen-
erations 500,001-1,000,000). The change in context from a
single-agent to a multi-agent environment has an instantaneous
effect on the evolution of agents, which can be seen in the
sharp increase in fitness around generation 500,000. This is
because agents can immediately capitalise on the changes in
environmental stimuli caused by the imperceptible actions of
the other agent in the environment – both with and without
neuromodulation. Neuromodulation is observed to help agents
to solve the multi-stage task in the RCD when they are
alone, and continue to help them adapt to their new, shared
environment when the context of the task is changed. The
benefit of neuromodulation is maintained for the remainder of
the evolutionary process, resulting in 99% of agents achieving
their goal, compared to only 73% of non-modulatory agents.
Activity-gating neuromodulation appears to increase both
the likelihood and the speed that agents evolve successful
solutions – both when they exist alone, and when they exist
together (Figure 4). This observation is more prominent when
agents evolve to solve a multi-stage task compared to a
single-stage task, thus showing that a simple change in task
complexity can affect the expected fitness of agents and their
ability to achieve their goals.
Similarly, a more obvious benefit arising from the use
of neuromodulation is seen in agents that evolve in multi-
agent environments than in those that evolve alone. This
is because the actions of each agent change the context
of the environment and therefore the state of each agent
within it, causing the variability of the environment to
increase. The benefit arising from a capacity for behavioural
plasticity through neuromodulation is therefore observed
to increase as the variability and unpredictability of the
task and environment increases. Agents that evolve to solve
multi-stage tasks in multi-agent environments are observed
to receive the highest benefit from immediate and reversible
behavioural changes in response to environmental stimuli.
Phenotypic plasticity is said to promote better adaptation to
variable and changing environmental conditions [3], which is
seen when evolving to solve multi-stage tasks in multi-agent
environments; these are the most dynamic and uncertain
conditions in this study.
The mean, median and variance for the best-in-population
fitness of each experiment was calculated after agents had
evolved for 500,000 generations, presented in Table III. This
analysis shows that neuromodulatory agents can expect a
higher median and mean fitness across all experiments; the
exception to this is when agents evolve to solve a single-stage
task together, in which case the median fitness is the same
with and without neuromodulation. Combined with the
results presented in Table II, neuromodulatory agents can
therefore not only be expected to have a higher mean and
median fitness, but they can be expected to solve the task and
achieve their goal more often than non-modulatory agents;
this is observed both in single- and multi-stage tasks, as
well as single- and multi-agent environments. The variance
in the best-in-population fitness after evolution is also lower
in neuromodulatory agents, which further exemplifies the
benefits of behavioural plasticity.
To analyse the effect that activity-gating neuromodulation
has on evolution further, Wilcoxon Signed Rank statistical tests
were conducted to compare the best-in-population fitnesses
of each experiment when agents evolve both without and
with neuromodulation. This is a non-parametric test used to
compare the medians of two paired distributions; the null
hypothesis of a two-tailed test is that the distribution medians
are equal, whereas one-tailed tests have the alternative hy-
pothesis that there is a directional difference in the distribution
medians (e.g. mn> mm). The null hypothesis can be rejected
when the calculated p-value is significant, below 0.05. These
Experiment Task Mean Median Variance
Alone S 0.58 0.5 0.00969697
Alone with NM S 0.67 0.7 0.005151515
Alone M 0.574 0.5 0.009418182
Alone with NM M 0.654 0.7 0.007155556
Together S 0.713 0.7 0.04215253
Together with NM S 0.836 0.7 0.02515556
Together M 0.754 0.7 0.04473131
Together with NM M 0.838 0.85 0.02864242
CE M 0.744 0.7 0.03844848
CE with NM M 0.852 0.95 0.02332929
Exp Task Statistical Test Alternative Hypothesis
(S/M) mn6=mmmn< mmmn> mm
Alone S 0.000000002588 0.000000001294 1
Together S 0.0002362 0.0001181 0.9999
Alone M 0.00000002994 0.00000001497 1
Together M 0.01594 0.00797 0.9922
CE M 0.0002593 0.0001296 0.9999
results are presented in Table IV. The two-tailed tests show
that there is a significant difference in median received fitness
between non-modulated and neuromodulated agents, for each
experiment in the study; the null hypothesis that the medians
of the two distributions are equal, can thus be rejected as
p < 0.05. Additionally, two one-tailed tests were conducted
to investigate whether there was a directional difference in
the distribution medians. These tests indicate that there is
a significant directional difference in the medians of the
two distributions, where the median of the non-modulatory
approach (mn) is lower than the modulatory approach (mm)
for each experiment conducted; furthermore, the contrasting
one-tailed test (mn> mm) shows no significant difference.
These results demonstrate that neuromodulation has both a
significantly different and a positive effect on the expected
fitness of agents, in all areas of the study.
Increasing environmental variability makes learning chal-
lenging for neural controllers, as encoded information must be
overwritten in order to learn new things when environmental
conditions change. The capacity to immediately and reversibly
change behaviour based on environmental stimuli is said to
promote adaptation in variable environments [3], [6]. We have
thus investigated the effect that activity-gating neuromodula-
tion has on an agent’s ability to evolve to succeed and to
make decisions in environments of increasing variability, by
exploring both single- and multi-stage tasks in single- and
multi-agent environments.
This study uses the River Crossing Dilemma [11] to explore
how agents evolve to solve multi-stage tasks; additionally we
propose a new adaptation of the testbed called the Protected
River Crossing Dilemma, in order to observe how agents
evolve to solve simpler single-stage tasks in less variable envi-
ronments. Our results demonstrate that behavioural plasticity
as a result of activity-gating neuromodulation has a signifi-
cant and increasingly positive effect on the expected fitness
of evolved agents, when the variability of the environment
increases; this behavioural plasticity is beneficial to create
adaptive agent controllers that can temporarily and reversibly
change behaviour in novel environments or situations.
We also show that when the context of an agent’s
environment changes from being individual to shared,
neuromodulation helps agents to adapt and succeed to the
new context and change in environmental stimuli. Often,
agents in this study will evolve an ‘exploitative’ fitness,
meaning that their success – and higher fitness – relies on the
actions of others in the environment; future work will explore
the extent to which behavioural plasticity enables agents to
maintain their goal-achieving behaviours when the presence of
other agents in the environment is unpredictable and uncertain.
This work was partially supported by the Research Council
of Norway through its Centres of Excellence scheme, project
number 262762. The authors would like to thank Aston
University and the University of Oslo for supporting the
research visits during this collaboration.
[1] T. Qian, T. F. Jaeger, and R. Aslin, “Learning to represent a
multi-context environment: More than detecting changes,Frontiers
in Psychology, vol. 3, p. 228, 2012. [Online]. Available: https:
[2] J. Yoder and L. Yaeger, “Evaluating topological models of neuromod-
ulation in polyworld,” in ALIFE 14: Proceedings of The Fourteenth
International Conference on the Synthesis and Simulation of Living
Systems, no. 26, 2014, pp. 916–923.
[3] M. Viney and A. Diaz, “Phenotypic plasticity in nematodes,Worm,
vol. 1, no. 2, pp. 98–106, 2012, pMID: 24058831. [Online]. Available:
[4] R. F. Oliveira, “Social plasticity in fish: integrating mechanisms and
function,” Journal of Fish Biology, vol. 81, no. 7, pp. 2127–2150, 2012.
[5] T. L. Rymer, N. Pillay, and C. Schradin, “Extinction or survival?
behavioral flexibility in response to environmental change in the african
striped mouse rhabdomys,” Sustainability, vol. 5, no. 1, pp. 163–186,
[6] E. C. Snell-Rood, “An overview of the evolutionary causes and conse-
quences of behavioural plasticity,Animal Behaviour, 2013.
[7] K. O. Stanley, J. Clune, J. Lehman, and R. Miikkulainen, “Designing
neural networks through neuroevolution,Nature Machine Intelligence,
[8] K. O. Ellefsen, J. B. Mouret, and J. Clune, “Neural Modularity Helps
Organisms Evolve to Learn New Skills without Forgetting Old Skills,”
PLoS Computational Biology, 2015.
[9] E. Robinson, T. Ellis, and A. Channon, “Neuroevolution of agents
capable of reactive and deliberative behaviours in novel and dynamic
environments,” in Advances in Artificial Life. Springer, 2007, pp. 1–
[10] J. Borg, A. Channon, and C. Day, “Discovering and Maintaining
Behaviours Inaccessible to Incremental Genetic Evolution Through
Transcription Errors and Cultural Transmission,” in ECAL, 2011.
[11] C. M. Barnes, A. Ek´
art, and P. R. Lewis, “Social action in socially situ-
ated agents,” in Proceedings of the IEEE 13th International Conference
on Self-Adaptive and Self-Organizing Systems, 2019, pp. 97–106.
[12] X. Yao, “Evolving artificial neural networks,” Proceedings of the IEEE,
vol. 87, no. 9, pp. 1423–1447, Sep. 1999.
[13] K. O. Stanley and R. Miikkulainen, “Evolving neural networks through
augmenting topologies,” Evolutionary Computation, vol. 10, no. 2, pp.
99–127, 2002.
[14] M. McCloskey and N. J. Cohen, “Catastrophic Interference in Con-
nectionist Networks: The Sequential Learning Problem,” Psychology of
Learning and Motivation - Advances in Research and Theory, 1989.
[15] J. A. Bullinaria, “Understanding the emergence of modularity in neural
systems,” Cognitive Science, 2007.
[16] R. Velez and J. Clune, “Diffusion-based neuromodulation can eliminate
catastrophic forgetting in simple neural networks,PLoS ONE, 2017.
[17] A. Dezfouli and B. W. Balleine, “Learning the structure of the world:
The adaptive nature of state-space and action representations in multi-
stage decision-making,” PLOS Computational Biology, vol. 15, no. 9,
pp. 1–22, 09 2019.
[18] L. F. Abbott, “Modulation of function and gated learning in a network
memory,Proceedings of the National Academy of Sciences of the
United States of America, 1990.
[19] N. Vecoven, D. Ernst, A. Wehenkel, and G. Drion, “Introducing neuro-
modulation in deep neural networks to learn adaptive behaviours,PLOS
ONE, vol. 15, no. 1, pp. 1–13, 01 2020.
[20] A. R. Daram, D. Kudithipudi, and A. Yanguas-Gil, “Task-based neu-
romodulation architecture for lifelong learning,” in 20th International
Symposium on Quality Electronic Design (ISQED), 2019, pp. 191–197.
[21] D. E. Asher, A. Zaldivar, B. Barton, A. A. Brewer, and J. L. Krichmar,
“Reciprocity and retaliation in social games with adaptive agents,IEEE
Transactions on Autonomous Mental Development, vol. 4, no. 3, pp.
226–238, 2012.
[22] F. Mery and J. G. Burns, “Behavioural plasticity: An interaction between
evolution and experience,Evolutionary Ecology, 2010.
[23] J. A. Stamps, “Individual differences in behavioural plasticities,Bio-
logical Reviews, 2016.
[24] A. W. Hamood and E. Marder, “Animal-to-animal variability in neuro-
modulation and circuit function,” in Cold Spring Harbor Symposia on
Quantitative Biology, vol. 79. Cold Spring Harbor Laboratory Press,
2014, pp. 21–28.
[25] A. Soltoggio, J. A. Bullinaria, C. Mattiussi, P. D¨
urr, and D. Floreano,
“Evolutionary advantages of neuromodulated plasticity in dynamic,
reward-based scenarios,” in Artificial Life XI: Proceedings of the 11th
International Conference on the Simulation and Synthesis of Living
Systems, ALIFE 2008, 2008.
[26] L. F. Abbott and S. B. Nelson, “Synaptic plasticity: taming the beast,
Nature Neuroscience, vol. 3, no. 11, pp. 1178–1183, 2000.
[27] W. S. Grant, J. Tanner, and L. Itti, “Biologically plausible learning in
neural networks with modulatory feedback,” Neural Networks, 2017.
[28] J. Huang, X. Ruan, N. Yu, Q. Fan, J. Li, and J. Cai, “A cognitive model
based on neuromodulated plasticity,” Computational Intelligence and
Neuroscience, 2016.
[29] J. L. Krichmar, “The neuromodulatory system: A framework for survival
and adaptive behavior in a challenging world,Adaptive Behavior,
vol. 16, no. 6, pp. 385–399, 2008.
[30] J. I. Espinosa-Ramos, E. Capecci, and N. Kasabov, “A computational
model of neuroreceptor-dependent plasticity (nrdp) based on spiking
neural networks,” IEEE Transactions on Cognitive and Developmental
Systems, vol. 11, no. 1, pp. 63–72, March 2019.
[31] N. Y. Masse, G. D. Grant, and D. J. Freedman, “Alleviating catastrophic
forgetting using context-dependent gating and synaptic stabilization,
Proceedings of the National Academy of Sciences, vol. 115, no. 44, pp.
E10 467–E10 475, 2018.
[32] S. Beaulieu, L. Frati, T. Miconi, J. Lehman, K. O. Stanley, J. Clune,
and N. Cheney, “Learning to Continually Learn,arXiv e-prints, p.
arXiv:2002.09571, Feb. 2020.
[33] P. E. Komers, “Behavioural plasticity in variable environments,” Cana-
dian Journal of Zoology, vol. 75, no. 2, pp. 161–169, 1997.
[34] K. Mogielski and T. Płatkowski, “A mechanism of dynamical interac-
tions for two-person social dilemmas,” Journal of Theoretical Biology,
[35] P. Kollock, “Social Dilemmas: The Anatomy of Cooperation,” Annual
Review of Sociology, 1998.
[36] J. Z. Leibo, V. Zambaldi, M. Lanctot, J. Marecki, and T. Graepel, “Multi-
agent Reinforcement Learning in Sequential Social Dilemmas,” Pro-
ceedings of the 16th Conference on Autonomous Agents and MultiAgent
Systems, 2017.
[37] P. A. van Lange, J. Joireman, C. D. Parks, and E. Van Dijk, “The
psychology of social dilemmas: A review,” Organizational Behavior and
Human Decision Processes, 2013.
[38] K. Sigmund and M. A. Nowak, “Evolutionary game theory,Current
Biology, 1999.
[39] S. X. Yang and M. Meng, “An efficient neural network method for
real-time motion planning with safety consideration,” Robotics and
Autonomous Systems, vol. 32, pp. 115–128, 2000.
[40] ——, “An efficient neural network approach to dynamic robot motion
planning,” Neural Networks, vol. 13, no. 2, pp. 143–148, 2000.
Full-text available
Animals excel at adapting their intentions, attention, and actions to the environment, making them remarkably efficient at interacting with a rich, unpredictable and ever-changing external world, a property that intelligent machines currently lack. Such an adaptation property relies heavily on cellular neuromodulation, the biological mechanism that dynamically controls intrinsic properties of neurons and their response to external stimuli in a context-dependent manner. In this paper, we take inspiration from cellular neuromodulation to construct a new deep neural network architecture that is specifically designed to learn adaptive behaviours. The network adaptation capabilities are tested on navigation benchmarks in a meta-reinforcement learning context and compared with state-of-the-art approaches. Results show that neuromodulation is capable of adapting an agent to different tasks and that neuromodulation-based approaches provide a promising way of improving adaptation of artificial systems.
Full-text available
State-space and action representations form the building blocks of decision-making processes in the brain; states map external cues to the current situation of the agent whereas actions provide the set of motor commands from which the agent can choose to achieve specific goals. Although these factors differ across environments, it is currently unknown whether or how accurately state and action representations are acquired by the agent because previous experiments have typically provided this information a priori through instruction or pre-training. Here we studied how state and action representations adapt to reflect the structure of the world when such a priori knowledge is not available. We used a sequential decision-making task in rats in which they were required to pass through multiple states before reaching the goal, and for which the number of states and how they map onto external cues were unknown a priori. We found that, early in training, animals selected actions as if the task was not sequential and outcomes were the immediate consequence of the most proximal action. During the course of training, however, rats recovered the true structure of the environment and made decisions based on the expanded state-space, reflecting the multiple stages of the task. Similarly, we found that the set of actions expanded with training, although the emergence of new action sequences was sensitive to the experimental parameters and specifics of the training procedure. We conclude that the profile of choices shows a gradual shift from simple representations to more complex structures compatible with the structure of the world.
Conference Paper
Full-text available
Two systems pursuing their own goals in a shared world can interact in ways that are not so explicit-such that the presence of another system alone can interfere with how one is able to achieve its own goals. Drawing inspiration from human psychology and the theory of social action, we propose the notion of employing social action in socially situated agents as a means of alleviating interference in interacting systems. Here we demonstrate that these specific issues of behavioural and evolutionary instability caused by the unintended consequences of interactions can be addressed with agents capable of a fusion of goal-rationality and traditional action, resulting in a stable society capable of achieving goals during the course of evolution.
Full-text available
Much of recent machine learning has focused on deep learning, in which neural network weights are trained through variants of stochastic gradient descent. An alternative approach comes from the field of neuroevolution, which harnesses evolutionary algorithms to optimize neural networks, inspired by the fact that natural brains themselves are the products of an evolutionary process. Neuroevolution enables important capabilities that are typically unavailable to gradient-based approaches, including learning neural network building blocks (for example activation functions), hyperparameters, architectures and even the algorithms for learning themselves. Neuroevolution also differs from deep learning (and deep reinforcement learning) by maintaining a population of solutions during search, enabling extreme exploration and massive parallelization. Finally, because neuroevolution research has (until recently) developed largely in isolation from gradient-based neural network research, it has developed many unique and effective techniques that should be effective in other machine learning areas too. This Review looks at several key aspects of modern neuroevolution, including large-scale computing, the benefits of novelty and diversity, the power of indirect encoding, and the field’s contributions to meta-learning and architecture search. Our hope is to inspire renewed interest in the field as it meets the potential of the increasing computation available today, to highlight how many of its ideas can provide an exciting resource for inspiration and hybridization to the deep learning, deep reinforcement learning and machine learning communities, and to explain how neuroevolution could prove to be a critical tool in the long-term pursuit of artificial general intelligence.
Full-text available
Humans and most animals can learn new tasks without forgetting old ones. However, training artificial neural networks (ANNs) on new tasks typically cause it to forget previously learned tasks. This phenomenon is the result of "catastrophic forgetting", in which training an ANN disrupts connection weights that were important for solving previous tasks, degrading task performance. Several recent studies have proposed methods to stabilize connection weights of ANNs that are deemed most important for solving a task, which helps alleviate catastrophic forgetting. Here, drawing inspiration from algorithms that are believed to be implemented in vivo, we propose a complementary method: adding a context-dependent gating signal, such that only sparse, mostly non-overlapping patterns of units are active for any one task. This method is easy to implement, requires little computational overhead, and allows ANNs to maintain high performance across large numbers of sequentially presented tasks when combined with weight stabilization. This work provides another example of how neuroscience-inspired algorithms can benefit ANN design and capability.
Full-text available
A long-term goal of AI is to produce agents that can learn a diversity of skills throughout their lifetimes and continuously improve those skills via experience. A longstanding obstacle towards that goal is catastrophic forgetting, which is when learning new information erases previously learned information. Catastrophic forgetting occurs in artificial neural networks (ANNs), which have fueled most recent advances in AI. A recent paper proposed that catastrophic forgetting in ANNs can be reduced by promoting modularity, which can limit forgetting by isolating task information to specific clusters of nodes and connections (functional modules). While the prior work did show that modular ANNs suffered less from catastrophic forgetting, it was not able to produce ANNs that possessed task-specific functional modules, thereby leaving the main theory regarding modularity and forgetting untested. We introduce diffusion-based neuromodulation, which simulates the release of diffusing, neuromodulatory chemicals within an ANN that can modulate (i.e. up or down regulate) learning in a spatial region. On the simple diagnostic problem from the prior work, diffusion-based neuromodulation 1) induces task-specific learning in groups of nodes and connections (task-specific localized learning), which 2) produces functional modules for each subtask, and 3) yields higher performance by eliminating catastrophic forgetting. Overall, our results suggest that diffusion-based neuromodulation promotes task-specific localized learning and functional modularity, which can help solve the challenging, but important problem of catastrophic forgetting.
Activity-dependent plasticity has attracted the interest of researchers for years in the domain of computational neuroscience, as the modification of synaptic efficacy occurs as a result of complex biochemical mechanisms that take place at a cellular level. In this paper, we introduce a phenomenological model implemented as an unsupervised learning rule for spiking neural networks (SNNs) based on the cross-talk between glutamatergic and ${\gamma }$ -aminobutyric acid (GABA)ergic neuroreceptors: ${N}$ -methyl- ${D}$ -aspartate receptor, ${\alpha }$ -amino-3-hydroxy-5-methyl-4-izoxazole-propionic acid receptors, GABA <sub xmlns:mml="" xmlns:xlink="">A</sub> , and GABA <sub xmlns:mml="" xmlns:xlink="">B</sub> . The proposed neuroreceptor-dependent plasticity (NRDP) model is implemented and demonstrated in an SNN environment, NeuCube, for modeling electroencephalography data. We show that the NRDP model can reproduce the generic spike-timing dependent plasticity behavior in an SNN. In addition, this can be used to simulate changes in excitatory/inhibitory balance in an SNN by altering neuroreceptors activity. More specifically, by varying the parameters that affect neuroreceptors activation, we can study how these changes would affect the learning and memory ability of a subject. In a therapeutic context, this makes it a promising tool for studying the regulatory mechanisms, where neuroreceptors cross-talk plays a crucial role. This can lead to new ways of early detection of neurological disorders and for better targeting drug treatments.
Conference Paper
Matrix games like Prisoner's Dilemma have guided research on social dilemmas for decades. However, they necessarily treat the choice to cooperate or defect as an atomic action. In real-world social dilemmas these choices are temporally extended. Cooperativeness is a property that applies to policies, not elementary actions. We introduce sequential social dilemmas that share the mixed incentive structure of matrix game social dilemmas but also require agents to learn policies that implement their strategic intentions. We analyze the dynamics of policies learned by multiple self-interested independent learning agents, each using its own deep Q-network, on two Markov games we introduce here: 1. a fruit Gathering game and 2. a Wolfpack hunting game. We characterize how learned behavior in each domain changes as a function of environmental factors including resource abundance. Our experiments show how conflict can emerge from competition over shared resources and shed light on how the sequential nature of real world social dilemmas affects cooperation.
Matrix games like Prisoner's Dilemma have guided research on social dilemmas for decades. However, they necessarily treat the choice to cooperate or defect as an atomic action. In real-world social dilemmas these choices are temporally extended. Cooperativeness is a property that applies to policies, not elementary actions. We introduce sequential social dilemmas that share the mixed incentive structure of matrix game social dilemmas but also require agents to learn policies that implement their strategic intentions. We analyze the dynamics of policies learned by multiple self-interested independent learning agents, each using its own deep Q-network, on two Markov games we introduce here: 1. a fruit Gathering game and 2. a Wolfpack hunting game. We characterize how learned behavior in each domain changes as a function of environmental factors including resource abundance. Our experiments show how conflict can emerge from competition over shared resources and shed light on how the sequential nature of real world social dilemmas affects cooperation.