When is an action caused from within? Quantifying the causal chain leading to actions in simulated agents

Abstract

An agent's actions can be influenced by external factors through the inputs it receives from the environment, as well as internal factors, such as memories or intrinsic preferences. The extent to which an agent's actions are "caused from within", as opposed to being externally driven, should depend on its sensor capacity as well as environmental demands for memory and context-dependent behavior. Here, we test this hypothesis using simulated agents ("animats"), equipped with small adaptive Markov Brains (MB) that evolve to solve a perceptual-categorization task under conditions varied with regards to the agents' sensor capacity and task difficulty. Using a novel formalism developed to identify and quantify the actual causes of occurrences ("what caused what?") in complex networks, we evaluate the direct causes of the animats' actions. In addition, we extend this framework to trace the causal chain ("causes of causes") leading to an animat's actions back in time, and compare the obtained spatio-temporal causal history across task conditions. We found that measures quantifying the extent to which an animat's actions are caused by internal factors (as opposed to being driven by the environment through its sensors) varied consistently with defining aspects of the task conditions they evolved to thrive in.

The Computational Boundary of a “Self”: Developmental Bioelectricity Drives Multicellularity and Scale-Free Cognition

“…What is the difference between a mechanism and decisionmaking (choice)? This has been widely discussed in the philosophy literature (Barandiaran and Moreno, 2006;Haig and Dennett, 2017) and is now becoming an important issue for bench biologists working in basal cognition of somatic cells and synthetic biology (Perkins and Swain, 2009;Balazsi et al, 2011;Reid et al, 2013Reid et al, , 2016Mitchell and Lim, 2016;Paul et al, 2016;Vesty et al, 2016;Bugaj et al, 2017) as well as for workers in artificial life (Juel et al, 2019). Without attempting to deal with this profound question in full, it can be mentioned that the perspective taken herein suggests that several factors contribute to a smooth transition between biochemical mechanism and agency: integration of remote events into the causal chain (spatial distance, and temporal distance -memory/ anticipation), and stochasticity (distance in terms of predictive capability).…”

Causal Composition: Structural Differences among Dynamically Equivalent Systems

“…This is demonstrated particularly well by recent computational studies involving simulated artificial agents with minimal cognitive architectures [ 15 , 23 , 68 , 69 ], whose behavior can easily be predicted. Yet, understanding what caused the agent to perform a particular action typically requires extensive additional analysis and cannot be addressed in purely reductionist or holistic terms [ 31 , 69 , 70 ].…”

Phi Fluctuates with Surprise: An empirical pre-study for the synthesis of the Free Energy Principle and Integrated Information Theory

“…To investigate the relation between surprisal, Φ and the state of the world external to a task-performing system, we did an evolutionary simulation study using the simulation framework MABE (Bohm, Lalejini, Schossau & Ofria, 2019). This framework, and the particular settings and environment implemented, have been used to study IIT in an evolutionary context before (Juel, Comolatti, Tononi & Albantakis, 2019). We attempted to replicate parts of the results from Albantakis et al (2014) where it was found that animats (artificial adaptive agents) on average evolved more concepts and higher values of Φ when the task environment was harder and more complex compared to when it was simpler.…”