Welcome to FoCAS Reading Room: Register or log in to tailor our content to your needs

FoCAS Reading Room is all about showing you what you're most interested in. Personalize the content on FoCAS Reading Room by choosing exactly what you want to see.

Laws of cognition for artificial intelligence

Identifying and studying the fundamental principles that underlie decision making could help us better understand and develop intelligent behaviour.

The endeavour to understand natural intelligence and reproduce it in artificial systems represents a relatively old scientific undertaking. However, many open questions remain. Several core philosophies dominate the discourse: the specific problems that require non-trivial treatment (e.g., speech recognition), which are addressed by well-developed and highly optimized algorithms put together for a particular purpose; algorithms that abstractly mimic the operation of biological designs (such as neural networks or revolutionary algorithms); and the detailed study of biological intelligence, on either a neural or higher-cognitive level.

Figure 1. An agent (blue) in a Minecraft-like world. The agent is building a bridge over a lava river to escape and incoming lava stream from the right. This is an example of an agent reshaping its environment in a creative way. Image credit: Christoph Salge.

Figure 1. An agent (blue) in a Minecraft-like world. The agent is building a bridge over a lava river to escape an incoming lava stream from the right. This is an example of an agent reshaping its environment in a creative way. (Image credit: Christoph Salge.)

These approaches, however, leave a number of issues unresolved. The engineering approach is increasingly successful in solving particular, well-delineated tasks to a high standard. However, it fails to shed light on the way in which biological organisms would solve such tasks, and leaves unclear how these approaches could be generalized beyond their narrow specification. For concrete applications, various niche-specific solutions must be patched together by a human engineer. Mimicking biological designs, on the other hand, does not determine to which extent the success of a particular bio-inspired algorithm occurs as a result of a fluke or because of an underlying universal principle. Furthermore, this approach does not make clear whether all success-relevant properties of the biological original have been captured in the computational abstraction. Finally, it is difficult to disentangle which component is essential to successful operation in biological intelligence and which is merely incidental to the evolutionary history of the investigated organism.

To address these issues, we follow cybernetic approaches, such as Ashby’s law of requisite variety.[1] We therefore use tools of information theory to formulate laws that govern cognition and decision-making. These laws are universal for any decision maker, whether that be a device or an organism, and are independent of implementation by a particular controller.[2]

This idea can be illustrated using an elementary example. Consider a robot that must reach a target location from a randomly selected starting point. In information-theoretic terms, this task can be interpreted as suppressing the uncertainty of the initial position to achieve the certainty of homing in on a particular target. In other words, the robot must reduce the complete initial entropy of its position to zero at the end of its run. To do so, the robot must necessarily process at least the same amount of information over the course of its trip as corresponds to the entropy of its initial location. The minimal information required for carrying out such a task is referred to as ‘relevant information,’ and is an invariant of the task. It is also independent of the strategy that the robot uses to navigate and make decisions. That is, whether the robot first captures its initial position precisely and then decides to move in a blind but accurate movement towards the target, or continuously acquires coarse heading information about the target and reactively converts it this information into movement, the total information processed never drops below the relevant information of the task. Cognitive architectures can therefore be evaluated exclusively with respect to information acquisition, processing and utilization without referring to details of the controller.

From these purely informational considerations, we can proceed to a consideration of the concrete navigational task in a particular environment and with a particular robotic design. These constraints on world and robot, and perception and action imprint a ‘signature’ on the information flowing through the robot. This signature depends only upon the system dynamics and the task, and (crucially) is invariant to the design of the robot controller. This allows us to gain significant insight into possible solutions and fundamental trade-offs between accuracy, processing costs and performance.

This formalism produces a number of expressive results and predictions. For instance, consider agents that are trying to maximize their future influence on the environment (given formally by the Shannon channel capacity between actuation and potential future states). This quantity (‘empowerment’) is an effective proxy for generating intrinsically motivated behaviour without externally imposed value functions.[3],[4] Its applications range from various balancing tasks without rewards to agents that build their own environment in creative ways. In other experiments, we assumed minimal working memory of the agents. Here, goal-directed behaviour gives rise to subgoals that emerge in a natural way, namely in locations where old information about the target action is superseded by new information.[5] These represent only two of many possible applications of this methodology.

Our information-theoretic methodology enriches the traditional study of cognitive processes by identifying laws of cognition and decision-making. It therefore complements the traditional lines of work (via observation-oriented experimental sciences or artefact-oriented engineered AI models) by using a rigorous framework that allows the discovery of the general quantitative organizational principles that govern the cognition of intelligent agents. In our future research, we aim to understand how information processing and cognitive processes can emerge from basic physics. In particular, we’re interested in understanding how these elements can ‘ascend’ towards higher-level capabilities,[6] an understanding of which is potentially linked with the development of hierarchies.


  1. H. Touchette and S. Lloyd, Information-Theoretic Limits of Control, Phys. Rev. Lett. 84, p. 1156, 2000.
  2. D. Polani, Information: currency of life?, HFSP J. 3, pp. 307–316, 2009.
  3. C. Salge, C. Glackin and D. Polani, Empowerment --- an introduction, Guided Self-Organization: Inception, pp. 67–114.
  4. T. Anthony, D. Polani and C. L. Nehaniv, 4. General self-motivation and strategy identification: case studies based on Sokoban and Pac-Man, IEEE Trans. Comput. Intell. AI in Games 6, pp. 1–17, 2013.
  5. S. G. van Dijk and D. Polani, Informational constraints-driven organization in goal-directed behavior, Adv. Complex Syst. 16, p. 1350016, 2013.
  6. A. C. Burgos and D. Polani, An informational study of the evolution of codes and of emerging concepts in populations of agents, Artif. Life 22, 2016.

About the Author

Add Your Commentary

To participate in discussion on FoCAS Reading Room you need to be logged in to your account.

Sign up for free or log in if you already have an account.

Related Articles

Cookies on FoCAS Reading Room website

Our website uses cookies to allow you to customise your content, use the site as a registered user, and to provide us with important information about visitors. By continuing to browse the site we'll assume that you are happy to receive all cookies set by FoCAS Reading Room.