Much of the recent neural network (NN) research has focused on the fields of ‘deep learning’ (DL), 
or ‘deep reinforcement learning’ (DRL)
. Neuroevolution (NE) and gradient-based learning methods have been used alternatively to train the weights of a NN. Metaphorically, the latter approach is inspired by learning-from-experience in which backpropagation
is the most common method to train the weights of a NN. This approach has proven to be impressively effective in supervised learning, especially in some reinforcement learning tasks so that there has been some strong claim that we can build human-intelligence by learning from scratch asblank-slates, without any innate knowledge , . Despite these impressive successes, there has been some objection to this line of thought that we are still very far from the so-called Artificial General Intelligence (AGI) and innate knowledge is necessary to reach AGI .
On another side, the NE approach uses a family of Evolutionary Algorithms which take inspiration from the fact that the natural brain is also the product of an evolutionary process. This approach was mostly developed in isolation from the gradient-based machine learning (ML) community, and has recently attracted interest from the ML community with several notable successes by OpenAI and Uber AI Labs .
Way back to the metaphorical level, in nature evolution and learning are not alternative but complementary. It has been said that intelligent behaviour produced by animal agents (including humans) is the product of the interplay between both nature and nurture, or innate and learned , , . Learning as a lifetime adaptation can change the evolutionary pathway and evolution can provide a base for learning to take place. Evolution can be thought of as not producing an absolute solution, but creating a learner (with or without brain) which then learn to find the solution more effectively than learning as blank-slates. This is greatly different from the way a number of EA researcher have adopted the idea of evolution to seek for a solution to optimization problems (including the one of optimizing neural networks).
This paper aims at two main things. First, I present a technique called evolving self-supervised neural networks, or neural networks that can teach themselves intrinsically without external supervisory signals. This network takes advantage of both evolutionary and gradient-descent learning to seek for its optimal weights. This is an important aspect of this contribution. Second, I simulate a multi-agent foraging world to test the performance of my proposed method.
Although there should possibly be a mixture of flavour, due to scope limit the reader is invited to have a look at  for an overview of neural network and deep learning, and  for a survey of neuroevolution approach to deep learning. In the remainder of this contribution, I initially present some prior research relating to learning and evolution in nature and in evolving neural networks. I then described the proposed technique and simulation undertaken. Results are then discussed, and concluded.
2 Related Work
The classic debate between nature vs nurture, or innate vs learned takes the separation between evolution vs learning in explaining behaviour , . Recent studies from Evolutionary Psychology ,  have shown that neither of these extremes can give the full account for the behaviour of living organisms, especially when it comes to explain the behaviour of humans. This is to say, our species are not totally blank-slates, and not totally instinctual too. Most of our adaptive behaviours are the product of the interaction between nature and nurture, or innate and learned, or evolution and learning.
There existed an intriguing idea called the Baldwin Effect presenting how a behaviour first learned can then become innate or partially innate in a Darwinian framework, not Lamarckian111Lamarckism says acquired behaviour is directly transmitted into the genome of the offspring, which has not been accepted in the mainstream of evolutionary biology. , . Interestingly, this idea was largely neglected in the mainstream of biology and psychology for over 100 years, and just gained more attention since the classic paper by the British Cognitive Scientist Geoffrey Hinton and his colleague entitled How Learning Can Guide Evolution in which an instance of the Baldwin Effect is demonstrated in a computer simulation . This paper stimulated a number of important follow-on studies including , , . All of these studies attempted to show that individual learning can promote the population-based search process of evolution.
in which the authors used a genetic algorithm to evolve the initial weights of a digit classifier neural network which then can be learned by backpropagation. They found that if the amount of learning is used properly, learning can take advantage of starting weights produced by evolution to further the classification performance.
Nolfi and his colleagues made a simulation of animats, or robots, controlled by neural networks situated in a grid-world, with discrete state and action spaces . Each agent lives in its own copy of the world, hence no mutual interaction. The evolutionary task is to evolve action strategies to collect food effectively, while each agent learns to predict the sensory inputs to neural networks for each time step. Learning was implemented using backpropagation based on the error between the actual and the predicted sensory inputs to update the weights of a neural network. It was shown that learning to predict can enhance the evolutionary search, hence increasing the performance of the robot.
Scientists at DeepMind recently also used the same idea of the Baldwin Effect, using a genetic algorithm to evolve the initial weights for deep neural networks . By combining the advantage of searching over a vast distribution of weights of evolutionary search and the exploitative power of a gradient descent learning, they reported a meta-learner that can solve a multiple tasks including regression and physical robot environments. This result is an indication to create meta-learning, another step towards AGI.
Generally learning in neural networks can be thought of as part of neural plasticity. There have been some other ideas, like evolving local learning rules to update the weights 
, evolution of neuromodulation which facilitates the information transfer between neurons in hopes of creating meta-learning. Please refer to  for more recent studies on evolving plastic neural networks. In short, most of the work use disembodied and unsituated neural networks in single-agent environment, having no mutual interaction as they solve their own problems, having no effect on other’s performance.
In this paper, I adopt a different method called evolving self-supervised neural networks (SSNNs) (which was presented in ). This technique differs from what has been presented so far in this section in that a SSNN consists of two separate network modules, namely action module and reinforcement module. The agent learns as the reinforcement module produces a signal to update the weights of the action module. This happens inside the agent whenever the agent moves and senses the environment, in need of no external supervision or reward, thus intrinsically motivated. This technique also differs greatly from traditional supervised learning in which a learning machine is provided with labels, and from reinforcement learning in which a reward function is set by the system engineer.
Unlike previous work, I simulate a situated multi-agent system. Autonomous multi-agent learning has been stated as a step toward AGI , . Situatedness here means the intelligent behaviour of an agent is coupled with its environment: the agent controls its own sensory inputs and actions in the environment. Please refer to  for a review of situated and embodied approach to cognitive science and AI – an inevitable part of general intelligence, including that of humans.
Our simulations are described in the following section. We shall be seeing how evolution and self-supervised learning interact to create more adaptive autonomous agents.
3.1 MiniWorld: Agents and Substances
Suppose that 20 agents situate in a continuous square 640x640 2D-world, called MiniWorld, which contains 60 substances of two types: food and poison. The multi-agent has to maximise its energy. Agents and substances are initially located at separate regions in MiniWorld depending on the world map. I propose 4 world maps (A, B, C, and D) in my simulations with details described in Figure 1222Please note that dim green lines are sketched only for the purpose of visualising the world map of the food and the agent, MiniWorld is a continuous world, not grid-like. One property of MiniWorld is it has no strict boundary, and I implement the so-called toroidal – this means when an agent moves beyond an edge, it appears in the opposite edge.. Each substance is represented by a square image of size 10x10. For simplicity I use a small binary bit sequence to represent the colour of the substance as described in Figure 1.
Each agent in MiniWorld has a squared body of size 10x10. When an agent’s body happens to collide with a substance, the substance is eaten and another piece of the same type randomly spawn in the same region but at a different location. The energy level of the agent increases or decreases by 1 if the agent eats food or toxin, respectively. The agent body somehow affects how the agent senses and acts in MiniWorld.
Every agent has a heading (in principle) of movement in the environment. To be more controllable, all agents are initialised with a horizontal heading (i.e. with 0 degree), not a random orientation. This deliberate setting is for purpose of our world maps. A visualisation of an agent and its relationship with substance in the environment is shown in Figure 2.
The default velocity (or speed) of each agent is 1. Every agent has 5 basic movements: Turn left by 20 degrees and move, move forward by double speed, turn right by 20 degrees and move, move in the reverse direction, and stop at the current location. For now, these rules are pre-defined by the system designer of MiniWorld 333We can imagine the perfect scenario like if an agent sees a food in front, it doubles the speed and move forward to catch the food. If the agent sees the food on the left (right), it should turn to the left (right) and move forward to the food particle. In cases in which the agent interprets that the substance is poisonous, it would be likely to turn and move in the reverse direction, or avoidance action. When an agent is not sure about its current circumstance, it may pause to think more.
Note that in our game, the agent has been equipped with very little domain knowledge about the world (e.g. state of the world, its current location or the relative distance between its current location to the source of substance). The agent is clearly autonomous. It has to experience the world on its own and acquire correct behaviour to find where the substance is. Moreover, the agent has no priori-knowledge about whether a substance it sees is food or poison. It has to implicitly develop an ability to discriminate between food and poison, through which it produces correct action (to approach or to avoid). Unlike reinforcement learning, no immediate reward is engineered in MiniWorld. The neural network method being proposed equips the agent with an intrinsic learning ability, which has been said a key to autonomous intelligence .
3.2 The Neural Network Controller
Each agent is controlled by a fully-connected neural network to determine its movements in the environment as shown in Figure 3
a). All neurons except the inputs use a sigmoidal activation function. All connections (or synaptic strengths) are initialised as Gaussian(0, 1). These weights are first initialised asinnate, but also have the potential to change during the lifetime of that agent. Please note that what an agent decides to do changes the world the agent lives in, changing the next sensory information it receives, hence the next behaviour. This forms a sensory-motor dynamics and a neural network acts as a situated cognitive module having the role to guide an agent to behave effectively and adaptively. This is situated cognition or situated intelligence .
In fact, the neural architecture as shown in Figure 3a) has no ability to supervise, or to teach itself. I extend this architecture to allow for self-supervised learning agents.
3.3 The Self-supervised Neural Architecture
To allow for self-supervision, the neural controller for each agent now has two modules, namely Action Module and Reinforcement Module. The action module is the same network as previously shown in Figure 3a). This module takes as sensory inputs and produce motor outputs. The reinforcement module has the same set of inputs as the action module, but has separate hidden and output neurons444For simplicity, I use the same topology for both modules. The goal of reinforcement network is to provide reinforcement signals to guide the behaviour of each agent through a self-supervised learning process. This architecture allows for an intrinsic learning process within the agent itself, no need for external supervision provided by the experimenter or reward like in DRL. Moreover, unlike actor-critic DRL the reinforcement module here does not provide reward signal and memory to update the action network, and it is not updated by changing the role of two networks . We shall see it is evolution that creates the self-supervision ability and the reinforcement signals, to make the evolved agent learns better. The description of the self-supervised neural network is visualised in Figure 3b). Simulations are descried below.
3.4 Simulation 1: Evolution alone (EVO)
I use an evolutionary algorithm to evolve the population of weight matrices, one of the most common approaches to neuroevolution .
The multi-agent system goes through a total of 100 generations, with 5000 time steps per generation. At each time step, an agent does the following activities: Perceiving MiniWorld through its sensors, computing its motor outputs by its neural network, moving in the environment which then updates its new heading and location. In evolution alone simulation, the agent cannot perform any kind of learning during its lifetime. After that, selection chooses individuals based on the number of food eaten in the foraging task used as the fitness value. The higher the number of food eaten, the higher the fitness value. For crossover, two individuals are selected as parents, namely and , in order to produce one offspring called . Crossover is implemented as the more successful parent has a higher the chance of passing its weights555Suppose ratio = /() (where , = the fitness of and , respectively). For each weight in the weight matrix of the child, we role a random number . If , w is copied from the same corresponding weight in ; otherwise from .
Once a child has been created, that child will be mutated based on a predefined mutation rate 0.05666If mutation occurs for each weight in the child, that weight is added by a random number from the range [-0.05, 0.05] – a slight mutation. After that, the newly born individual is placed in a new population. This process is repeated until the whole new 20 agents are created.
3.5 Simulation 2: Evolving Self-supervised agents
In this simulation, I allow lifetime learning, in addition to the evolutionary algorithm, to update the weights of neural network controllers. I evolve a population of Self-supervised agents – agents that can supervise, or teach themselves. During the lifetime of an agent, the reinforcement modules produce outputs in order to guide the weight-updating process of the action module. Only the weights of action modules can be changed by learning, the weights of reinforcement module are genetically specified in the same evolutionary process as specified above in Evolution alone.
I use the same parameter setting for evolution as in EVO simulation above. At each time step, an agent does the following activities: Receiving sensory inputs, computing its output, updating its new heading and location, and updating the weights in action module by self-teaching. After 5000 steps per generation, the multi-agent system undergoes selection and reproduction processes as in Evolution alone777Remember that we are fitting self-learning and evolution in a Darwinian framework, not Lamarckian. This means what will be learned during the lifetime of an agent (the weights in action module) is not directly passed down onto the offspring.
3.6 Simulation 3: Self-supervised agents Alone
I conduct another simulation in which all agents are self-taught agents – having self-taught networks that can teach themselves during lifetime. What differs from simulation 2 is that at the beginning of every generation, all weights are randomly initialised, rather than updated by an evolutionary algorithm like in simulation 1. The learning agents here are initialised as blank-slates, or tabula rasa, having no predisposition or some sort of priori knowledge about the world before learning. We are curious whether evolution brings any benefit to learning in MiniWorld.
4 Results, Analysis, and Discussion
We compare the best and the average energy (fitness) absorbed by the multi-agent system in MiniWorld. All results are averaged over 30 independent runs.
Before going into the result, through the visualisation of world maps and the description of what the agent has equipped, we can arrange world map A-B-C-D in the increasing order of difficulty to forage. In map A, all agents are initially born with a more tendency toward the food. The scenario is reversed in map B when the agent is born with a more tendency toward the poison. Map C and map D may present an even more difficulty for our game since the agent is born with no tendency to experience anything in the world888This is an important point because as described above, our agents are provided little information about the world, and they are autonomous and situated – what they sense can affect their movement in the world, which then affect their future behaviour, so on and so on. Without good movement they may not able to sense more relevant information which then helps produce more relevant movement in the world.. In fact, to be effective the agent has to learn two tasks: to sense the world correctly to find the substance, and from that develop the ability to discriminate between food and poison, which will be used to guide its movements.
4.1 Self-supervised Learning Facilitates Evolution
First we would like to see how self-supervised learning facilitates the evolutionary search by comparing EVO+Self-supervised (EVO+SS) and EVO alone. A similar trend can be observed in nearly all experimental settings is that there is a decrease in performance from A to B to C to D. This is understandable because of the increasing difficulty to forage in these maps as just analysed above.
How each type of experimental setup performs compared to each other? It can be seen that EVO+SS outperforms EVO alone in all maps with respect to both the best and the average energy. Specifically, in the easiest case (map A) we see that the best agent in EVO+SS, on average, is around 60 energy higher the best agent in EVO alone. Overall, the whole EVO+SS system absorbs 800 (40 on average) energy higher than the EVO alone. In map B, a harder case, we that the difference between these two systems (EVO and EVO+SS) is reduced, yet EVO+SS still has a clearly better performance.
Interestingly, in map C and D the EVO alone system cannot eat at all. This is predictable. Simply speaking, an agent that cannot learn can only use its innate ability, hardwired in its brain, to search the environment. However, as analysed above, the agent in EVO alone is born without any tendency to sense relevant information (about the substance) in the environment, and also has no ability to change its motor program hardwired in its brain. This is why they cannot eat anything at all since their actions are repeatedly making no sense in MiniWorld. Conversely, the system of self-supervised agents still can absorb energy in maps C and D.
Moreover, the boxplots in Figure 6 again show that EVO+SS outperforms EVO alone in all setups, and that this difference is statistically significant.
Not just the performance, one curious question arising here is why self-supervised agents can behave that way? One plausible explanation for this is the effect of self-supervised learning on evolution: By doing self-supervised learning, the EVO+SS system can facilitate the evolutionary search. This dominance helps show that the agent in EVO+SS may have successfully developed the ability to sense the world correctly and the ability to discriminate between food and poison, so that they can eat more food and less poison.
4.2 Evolution Facilitates Self-supervised Learning
We conduct the third simulation in which the neural networks controlling self-supervised agents are all randomly initialised at the beginning of each generation, without the participation of evolution. It can be observed in Figure 4 and Figure 5 that in all maps, the Self-supervised alone (SS) system has a clearly lower performance than EVO+SS, especially when it comes to the harder cases (map C and map D). The difference is statistically significant as shown in Figure 6. It is also interesting that in our simulations, the best blank-slate SS agent cannot outperform the best in EVO alone in easier cases (map A and B), but has some small advantage in harder cases (map C and D) when the agent in EVO alone cannot search for any food at all.
Another thing to be seen here is the average energy of the SS system is relatively close to 0 while the best agent still can absorb a positive amount of energy in all maps. This behaviour is interesting because it shows that through self-supervised learning some agent can develop a relatively effective behaviour to both explore the world ((even if it is born with tendency to reach nothing in map C and D), and behave as if it can eat more food than poison. This also indicates that the whole SS system without evolution cannot successfully develop the ability to discriminate between food and poison, hence the average energy is zero.
This can be explained as with the ability to teach oneself by leveraging the difference between the action and the reinforcement modules, the weights of the action module of some agent may have been changed to produce better movements. The agent that can reach the food source has a higher chance of being selected to leave offspring. Thus, its good genetic information, consisting of the initial weights of both the action and the reinforcement modules, is more likely to proliferate. When an agent is selected for reproduction, its self-supervising, or self-teaching, ability is likely to be also promoted in later generations. It is this ability that has made future evolved self-supervised agents better at teaching themselves in order to develop more effective movement in MiniWorld. This process repeats as what has given advantage during lifetime of the self-taught agent is preserved and promoted by the evolutionary process.
It is plausible conclude that some sort-of intelligent foraging behaviour has emerged in our autonomous multi-agent system, through the interaction between evolution and self-learning, and this is what I have observed in my simulation999The video demonstration has been made anonymously in this link https://youtu.be/aUlLJfisAHA.. Through self-taught learning, some initial advantage behaviour would have been emerged, and that behaviour changed the evolutionary pathway, which further that adaptive behaviour in the future. This is, through the evolutionary process, some priori-knowledge about the environment can be encoded in the neural network controller. This makes agents having predisposition to learn adaptive behaviours learn more effectively than blank-slate agents. This is the interplay between evolution and self-supervised learning.
5 Conclusion and Further Discussion
I have presented and analysed a technique called Evolving Self-supervised Neural Networks. Behind this technique is the interplay between evolution and learning in the form of self-supervision in order to produce ever more adaptive and intelligent behaviour. This proposed network has been used to control a situated autonomous multi-agent system in which the agent requires very little domain and human-engineered knowledge. Experimental results in increasingly difficult maps have shown that the Evolved Self-supervised system outperforms both the EVO and the Self-supervised learning system alone. The interplay between self-learning and evolution has also been demonstrated in the sense that evolution facilitates future self-supervised learning, better than learning as blank-slates. Through this interaction, an there is an emergence of intelligent behaviour in our system.
Here I stress more onto the philosophical idea that self-learning and evolution can interact and benefit each other in evolving intelligence. This differs from traditional point of view in EAs, and also the blank-slate point of view in that here we have evolved a learner better at learning to solve a problem with more innate information gradually encoded by evolution, rather than having evolved a solution to that problem. This fits with the current understanding of human cognition and intelligent behaviour in cognitive science and evolutionary epistemology. It seems to open another way to reach human-like intelligence in the computer: Not evolve to solve, not learn as blank-slate, but evolve a learner to solve.
The computational method is simple enough to illustrate the idea, but still has some indications. The idea of self-taught neural networks can be powerful when there is no external supervision (or label provided from external data). This opens a way to create intelligent autonomous multi-agent learning, in hopes of producing autonomous intelligence , 
. The algorithm and technique used in this paper can also be a potential technique to solve unsupervised learning, or learning with limited label data (weak supervision), especially in reinforcement learning and games . We are curious whether evolution can provide a better base to learn than learning as blank-slates like what was claimed by DeepMind in games . Indeed, the shallow network used in this paper does not restrict the application of the core philosophical idea into deep neural networks, as long as we can combine evolutionary search and the idea of self-taught neural architecture by employing variants of gradient-based learning.
There is some limitation that should not be neglected, including the use of a fixed neural architecture. This limitation has also been raised in a few in current deep learning research , . One plausible solution could be evolving both the weights and the topology of a neural network , . This is an interesting pathway for future work if we can evolve variable self-supervised neural architecture which can be an intrinsically general neural learner.
-  Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. Nature, 521:436 EP –, 05 2015.
-  Jürgen Schmidhuber. Deep learning in neural networks: An overview. CoRR, abs/1404.7828, 2014.
-  Volodymyr Mnih, Koray Kavukcuoglu, David Silver, and Andrei A. Rusu. Human-level control through deep reinforcement learning. Nature, 518(7540):529–533, feb 2015.
-  David E. Rumelhart, Geoffrey E. Hinton, and Ronald J. Williams. Learning representations by back-propagating errors. Nature, 323(6088):533–536, oct 1986.
-  David Silver, Julian Schrittwieser, Karen Simonyan, Ioannis Antonoglou, and Aja Huang. Mastering the game of go without human knowledge. Nature, 550(7676):354–359, oct 2017.
-  Gary Marcus. Innateness, alphazero, and artificial intelligence. CoRR, abs/1801.05667, 2018.
-  Tim Salimans, Jonathan Ho, Xi Chen, and Ilya Sutskever. Evolution strategies as a scalable alternative to reinforcement learning. CoRR, abs/1703.03864, 2017.
-  Felipe Petroski Such, Vashisht Madhavan, Edoardo Conti, Joel Lehman, Kenneth O. Stanley, and Jeff Clune. Deep neuroevolution: Genetic algorithms are a competitive alternative for training deep neural networks for reinforcement learning. CoRR, abs/1712.06567, 2017.
-  S Pinker. The Blank Slate (2002/2016). New York, NY: Viking, 2016.
-  Joni Y. Sasaki and Heejung S. Kim. Nature, nurture, and their interplay. Journal of Cross-Cultural Psychology, 48(1):4–22, nov 2016.
-  Richard Samuels. Innateness in cognitive science. Trends in Cognitive Sciences, 8(3):136 – 141, 2004.
-  Kenneth O. Stanley, Jeff Clune, Joel Lehman, and Risto Miikkulainen. Designing neural networks through neuroevolution. Nature Machine Intelligence, 1(1):24–35, jan 2019.
-  J.H. Barkow, L. Cosmides, and J. Tooby. The Adapted Mind: Evolutionary Psychology and the Generation of Culture. Oxford University Press, 1995.
-  S Pinker. How the Mind Works (1997/2009). New York, NY: W. W. Norton & Company, 2009 edition, 2009.
-  J. Mark Baldwin. A new factor in evolution. The American Naturalist, 30(354):441–451, 1896.
-  George Gaylord Simpson. The baldwin effect. Evolution, 7(2):110, jun 1953.
-  Geoffrey E. Hinton and Steven J. Nowlan. How learning can guide evolution. Complex Systems, 1:495–502, 1987.
-  Stefano Nolfi, Domenico Parisi, and Jeffrey L. Elman. Learning and evolution in neural networks. Adaptive Behavior, 3(1):5–28, 1994.
-  Giles Mayley. Guiding or hiding: Explorations into the effects of learning on the rate of evolution. In In Proceedings of the Fourth European Conference on Artificial Life, pages 135–144. MIT Press, 1997.
-  Inman Harvey. Is there another new factor in evolution? Evol. Comput., 4(3):313–329, September 1996.
-  Ron Keesing and David G. Stork. Evolution and learning in neural networks: The number and distribution of learning trials affect the rate of evolution. In NIPS 1990, 1990.
-  Chrisantha Thomas Fernando, Jakub Sygnowski, Simon Osindero, Jane Wang, Tom Schaul, Denis Teplyashin, Pablo Sprechmann, Alexander Pritzel, and Andrei A. Rusu. Meta-learning by the baldwin effect. CoRR, abs/1806.07917, 2018.
-  Samy Bengio, Yoshua Bengio, Jocelyn Cloutier, and Jan Gecsei. On the optimization of a synaptic learning rule. In D. S. Levine and W. R. Elsberry, editors, Optimality in Biological and Artificial Networks. Lawrence Erlbaum, 1995.
-  Kenji Doya. Metalearning and neuromodulation. Neural Networks, 15(4):495 – 506, 2002.
-  Andrea Soltoggio, Kenneth O. Stanley, and Sebastian Risi. Born to learn: The inspiration, progress, and future of evolved plastic artificial neural networks. Neural Networks, 108:48–67, dec 2018.
-  Nam Le. Evolving self-taught neural networks: The baldwin effect and the emergence of intelligence. In 10th Symposium on AI & Games, in AISB Convention 2019, Falmouth, UK, 16-18 Apri 2019.
-  Peter Stone. Learning and multiagent reasoning for autonomous agents. In Proceedings of the 20th International Joint Conference on Artifical Intelligence, IJCAI’07, pages 13–30, San Francisco, CA, USA, 2007. Morgan Kaufmann Publishers Inc.
-  Ryan Lowe, Yi Wu, Aviv Tamar, Jean Harb, Pieter Abbeel, and Igor Mordatch. Multi-agent actor-critic for mixed cooperative-competitive environments. CoRR, abs/1706.02275, 2017.
-  Michael L. Anderson. Embodied cognition: A field guide. Artificial Intelligence, 149(1):91 – 130, 2003.
-  Nuttapong Chentanez, Andrew G. Barto, and Satinder P. Singh. Intrinsically motivated reinforcement learning. In L. K. Saul, Y. Weiss, and L. Bottou, editors, Advances in Neural Information Processing Systems 17, pages 1281–1288. MIT Press, 2005.
-  Zhi-Hua Zhou. A brief introduction to weakly supervised learning. National Science Review, 5(1):44–53, aug 2017.
-  R.S. Sutton, A.G. Barto, and F. Bach. Reinforcement Learning: An Introduction. MIT Press, 2018.
-  Kenneth O. Stanley and Risto Miikkulainen. Evolving neural networks through augmenting topologies. Evol. Comput., 10(2):99–127, June 2002.