Active Inference: Demystified and Compared
Open accessThe DOI matches its Crossref record; 31 of the 31 references it cites were found in a registry or an index.
Abstract
Active inference is a first principle account of how autonomous agents operate in dynamic, nonstationary environments. This problem is also considered in reinforcement learning, but limited work exists on comparing the two approaches on the same discrete-state environments. In this letter, we provide (1) an accessible overview of the discrete-state formulation of active inference, highlighting natural behaviors in active inference that are generally engineered in reinforcement learning, and (2) an explicit discrete-state comparison between active inference and reinforcement learning on an OpenAI gym baseline. We begin by providing a condensed overview of the active inference literature, in particular viewing the various natural behaviors of active inference agents through the lens of reinforcement learning. We show that by operating in a pure belief-based setting, active inference agents can carry out epistemic exploration—and account for uncertainty about their environment—in a Bayes-optimal fashion. Furthermore, we show that the reliance on an explicit reward signal in reinforcement learning is removed in active inference, where reward can simply be treated as another observation we have a preference over; even in the total absence of rewards, agent behaviors are learned through preference learning. We make these properties explicit by showing two scenarios in which active inference agents can infer behaviors in reward-free environments compared to both Q-learning and Bayesian model-based reinforcement learning agents and by placing zero prior preferences over rewards and learning the prior preferences over the observations corresponding to reward. We conclude by noting that this formalism can be applied to more complex settings (e.g., robotic arm movement, Atari games) if appropriate generative models can be formulated. In short, we aim to demystify the behavior of active inference agents by presenting an accessible discrete state-space and time formulation and demonstrate these behaviors in a OpenAI gym environment, alongside reinforcement learning agents.
Overview
The claims and statistics of this paper haven’t been checked yet.
Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
The claims and statistics of this paper haven’t been checked yet.
Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
The self-check needs the paper’s full text.
Citations
Cites
Watkins, C. J. C. H., & Dayan, P. (1992) · Machine Learning · doi:10.1007/bf00992698 · Cited by 9,083
Asynchronous Methods for Deep Reinforcement Learning
Mnih, V., Badia, A. P., Mirza, M., Graves, A., Lillicrap, T. P., Harley, T., Silver, D., & Kavukcuoglu, K. (2016) · arXiv · doi:10.48550/arxiv.1602.01783 · Cited by 7,132 · cited 1×
“Many different algorithms have been used to address this, including (cid:2) -greedy (Vermorel & Mohri, 2005; Mnih et al., 2013, 2016), action selection based on action-utility (Sutton, 1990) and counter-based strategies (Wiering & Schmidhuber, 1998; Tijsma, Drugan, & Wiering, 2016).”
MethodSoft actor-critic Off-policy maximum entropy deep reinforcement learning with a stochastic actor
Haarnoja, T., Zhou, A., Abbeel, P., & Levine, S. (2018) · arXiv · doi:10.48550/arxiv.1801.01290 · Cited by 5,594 · cited 1×
“Having said this, sophisticated formulations of reinforcement learning, such as soft actor-critic (Haarnoja, Zhou, Abbeel, & Levine, 2018; Cesa-Bianchi et al., 2017), have addressed this issue.”
Stochastic thermodynamics, fluctuation theorems and molecular machines
Seifert, U. (2012) · Reports on Progress in Physics · doi:10.1088/0034-4885/75/12/126001 · Cited by 3,300 · cited 2×
“of non-equilibrium physics, it is analogous to self-organization or selfassembly (Crauel & Flandoli, 1994; Seifert, 2012; Friston, 2019).”
Curiosity-Driven Exploration by Self-Supervised Prediction
Pathak, D., Agrawal, P., Efros, A. A., & Darrell, T. (2017) · 2017 IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) · doi:10.1109/cvprw.2017.70 · Cited by 1,815 · cited 3×
“This is in contrast to vanilla reinforcement learning, where it is less clear how to naturally account for learning an intrinsic reward function, with many competing approaches (Still & Precup, 2012; Mohamed & Rezende, 2015; Pathak et al., 2017).”
BackgroundConcrete Problems in AI Safety
Amodei, D., Olah, C., Steinhardt, J., Christiano, P., Schulman, J., & Mané, D. (2016) · arXiv · doi:10.48550/arxiv.1606.06565 · Cited by 1,314 · cited 1×
“Consequently, crafting appropriate reward functions is not easy, and it is possible for agents to learn suboptimal actions if the reward function is poorly specified (Amodei et al., 2016).”
BackgroundActive inference: a process theory
Friston, K., FitzGerald, T., Rigoli, F., Schwartenbeck, P., & Pezzulo, G. (2017) · Neural Computation · doi:10.1162/neco_a_00912 · Cited by 1,156 · cited 9×
“…emerges naturally from active inference, where the generative model determines the nature of discounting (based on parameter-capturing precision), with predictions in the distant future being less precise and thus discounted in a Bayes-optimal fashion (Friston, FitzGerald, et al., 2017).”
BackgroundMethodInfluentialSelf-citationOn the theory of dynamic programming
Bellman, R. (1952) · Proceedings of the National Academy of Sciences · doi:10.1073/pnas.38.8.716 · Cited by 998
Attractors for random dynamical systems
Crauel, H., & Flandoli, F. (1994) · Probability Theory and Related Fields · doi:10.1007/bf01193705 · Cited by 985
Optimal control of Markov processes with incomplete state information
Åström, K. J. (1965) · Journal of Mathematical Analysis and Applications · doi:10.1016/0022-247x(65)90154-x · Cited by 854 · cited 5×
“This allows the problem to be framed as a partially observable Markov decision process (POMDP) (Astrom, 1965), where the generative model allows us to make inferences about true states given outcomes.”
BackgroundInfluentialActive inference and epistemic value
Friston, K., Rigoli, F., Ognibene, D., Mathys, C., Fitzgerald, T., & Pezzulo, G. (2015) · Cognitive Neuroscience · doi:10.1080/17588928.2015.1020053 · Cited by 731
Self-citationFriston, K., FitzGerald, T., Rigoli, F., Schwartenbeck, P., ODoherty, J., & Pezzulo, G. (2016) · Neuroscience & Biobehavioral Reviews · doi:10.1016/j.neubiorev.2016.06.022 · Cited by 674
Self-citationAction understanding and active inference
Friston, K., Mattout, J., & Kilner, J. (2011) · Biological Cybernetics · doi:10.1007/s00422-011-0424-z · Cited by 665
Self-citationReinforcement learning or active inference?
Friston, K. J., Daunizeau, J., & Kiebel, S. J. (2009) · PLoS ONE · doi:10.1371/journal.pone.0006421 · Cited by 443
Self-citationThe graphical brain: Belief propagation and active inference
Friston, K. J., Parr, T., & de Vries, B. (2017) · Network Neuroscience · doi:10.1162/netn_a_00018 · Cited by 415 · cited 4×
“…2017; Friston, Parr, et al., 2017; Parr & Friston, 2017), continuous state-space models (Buckley, Kim, McGregor, & Seth, 2017; Parr & Friston, 2019a; Ueltzhöffer, 2018) or mixed models with both discrete and continuous states as described in Friston, Parr, et al. (2017) and Parr and Friston (2018).”
BackgroundInfluentialSelf-citationReinforcement Learning and Control as Probabilistic Inference: Tutorial and Review
Levine, S. (2018) · arXiv · doi:10.48550/arxiv.1805.00909 · Cited by 366 · cited 3×
“To make meaningful comparisons between active inference and reinforcement learning, we commit to the definition of reinforcement learning in Sutton and Barto (1998, 2018): “Reinforcement learning is learning what to do—how to map situations to actions—so as to maximize a numerical reward signal.””
BackgroundMethodLearning latent structure: carving nature at its joints
Gershman, S. J., & Niv, Y. (2010) · Current Opinion in Neurobiology · doi:10.1016/j.conb.2010.02.008 · Cited by 326
The anatomy of choice: dopamine and decision-making.
Friston, K., Schwartenbeck, P., FitzGerald, T., Moutoussis, M., Behrens, T., & Dolan, R. J. (2014) · Philosophical Transactions of the Royal Society B: Biological Sciences · doi:10.1098/rstb.2013.0481 · Cited by 289
Self-citationA tutorial on the free-energy framework for modelling perception and learning
Bogacz, R. (2017) · Journal of Mathematical Psychology · doi:10.1016/j.jmp.2015.11.003 · Cited by 283 · cited 2×
“It postulates that agents maintain homeostasis by residing in (attracting) states that minimize surprise (Friston, Mattout, & Kilner, 2011; Bogacz, 2017).”
BackgroundMarkov decision processes: a tool for sequential decision making under uncertainty
Alagoz, O., Hsu, H., Schaefer, A. J., & Roberts, M. S. (2010) · Medical Decision Making · doi:10.1177/0272989x09353194 · Cited by 266
Uncertainty, epistemics and active inference
Parr, T., & Friston, K. J. (2017) · Journal of The Royal Society Interface · doi:10.1098/rsif.2017.0376 · Cited by 252
Self-citationActive inference and agency: optimal control without cost functions
Friston, K., Samothrakis, S., & Montague, R. (2012) · Biological Cybernetics · doi:10.1007/s00422-012-0512-8 · Cited by 241 · cited 1×
“Normatively speaking, active inference dispenses with the Bellman optimality principle and replaces it with a (variational) principle of least action (see Friston, Samothrakis, & Montague, 2012, for further discussion).”
MethodSelf-citationGeneralised free energy and active inference
Parr, T., & Friston, K. J. (2019) · Biological Cybernetics · doi:10.1007/s00422-019-00805-w · Cited by 236 · cited 5×
“…FitzGerald, et al., 2017; Friston, Parr, et al., 2017; Parr & Friston, 2017), continuous state-space models (Buckley, Kim, McGregor, & Seth, 2017; Parr & Friston, 2019a; Ueltzhöffer, 2018) or mixed models with both discrete and continuous states as described in Friston, Parr, et al. (2017) and…”
BackgroundInfluentialSelf-citationA free energy principle for a particular physics
Friston, K. (2019) · arXiv · doi:10.48550/arxiv.1906.10184 · Cited by 222 · cited 10×
“…et al., 2017; Friston, Parr, et al., 2017; Parr & Friston, 2017), continuous state-space models (Buckley, Kim, McGregor, & Seth, 2017; Parr & Friston, 2019a; Ueltzhöffer, 2018) or mixed models with both discrete and continuous states as described in Friston, Parr, et al. (2017) and Parr and…”
BackgroundMethodInfluentialSelf-citationRepresentation and timing in theories of the dopamine system
Daw, N. D., Courville, A. C., & Touretzky, D. S. (2006) · Neural Computation · doi:10.1162/neco.2006.18.7.1637 · Cited by 173
An information-theoretic approach to curiosity-driven reinforcement learning
Still, S., & Precup, D. (2012) · Theory in Biosciences · doi:10.1007/s12064-011-0142-z · Cited by 165 · cited 2×
“Information-theoretic approaches have also been explored in a reinforcement learning context (Still & Precup, 2012; Mohamed & Rezende, 2015; Blau, Ott, & Ramos, 2019).”
BackgroundScene Construction, Visual Foraging, and Active Inference
Mirza, M. B., Adams, R. A., Mathys, C. D., & Friston, K. J. (2016) · Frontiers in Computational Neuroscience · doi:10.3389/fncom.2016.00056 · Cited by 165
Self-citationDiscovering latent causes in reinforcement learning
Gershman, S. J., Norman, K. A., & Niv, Y. (2015) · Current Opinion in Behavioral Sciences · doi:10.1016/j.cobeha.2015.07.007 · Cited by 151
Neuronal message passing using mean-field, bethe, and marginal approximations
Parr, T., Markovic, D., Kiebel, S. J., & Friston, K. J. (2019) · Scientific Reports · doi:10.1038/s41598-018-38246-3 · Cited by 128
Self-citationToward the neural implementation of structure learning
Tervo, D. G. R., Tenenbaum, J. B., & Gershman, S. J. (2016) · Current Opinion in Neurobiology · doi:10.1016/j.conb.2016.01.014 · Cited by 106
Ueltzhöffer, K. (2018) · Biological Cybernetics · doi:10.1007/s00422-018-0785-7 · Cited by 98 · cited 2×
“…2017; Friston, Parr, et al., 2017; Parr & Friston, 2017), continuous state-space models (Buckley, Kim, McGregor, & Seth, 2017; Parr & Friston, 2019a; Ueltzhöffer, 2018) or mixed models with both discrete and continuous states as described in Friston, Parr, et al. (2017) and Parr and Friston (2018).”
Background
Research works printed but not linked: 60
- Al-Shedivat M., 2017, Continuous adaptation via meta-learning in nonstationary and competitive environments
- Attias H., 2003, AISTATS
- 10.1016/B978-0-12-398532-3.00007-5 · DOI
- Beal M. J., 2003, Variational algorithms for approximate Bayesian inference
- Beck J., 2012, Advances in neural information processing systems, 25, 3059
- Blau T., 2019, Bayesian curiosity for efficient exploration in reinforcement learning.
- 10.1080/01621459.2017.1285773 · DOI · Paper page
- 10.1016/j.tics.2012.08.006 · DOI · Paper page
- 10.1016/j.jmp.2017.09.004 · DOI · Paper page
- Burda Y., 2018, Exploration by random network distillation
- Camacho E. F., 2013, Model predictive control
- Cao F., 2012, Advances in neural information processing systems, 25, 73
- Cesa-Bianchi N., 2017, Advances in neural information processing systems, 25, 6284
- 10.1073/pnas.1512144113 · DOI · Paper page
- 10.1016/j.bpsc.2018.06.010 · DOI · Paper page
- Da Costa L., 2020, Active inference on discrete state-spaces: A synthesis
- 10.1016/j.neuron.2011.02.027 · DOI · Paper page
- Dearden R., 2013, Model-based Bayesian exploration
- Dearden R., 1998, Proceedings of the 15th National Conference on Artificial Intelligence, 761
- Foerster J. N., 2017, Learning with opponent-learning awareness
- 10.1016/j.neubiorev.2017.04.009 · DOI · Paper page
- 10.1162/neco.2007.19.12.3173 · DOI · Paper page
- 10.1007/s10994-012-5313-8 · DOI
- 10.1002/9781119159193.ch33 · DOI · Paper page
- 10.1146/annurev-psych-122414-033625 · DOI · Paper page
- 10.1561/2200000049 · DOI · Paper page
- Igl M., 2019, Advances in neural information processing systems, 32, 13978
- Igl M., 2018, Deep variational reinforcement learning for POMDPs
- Kolter J. Z., 2009, Proceedings of the 26th Annual International Conference on Machine Learning, 513, 10.1145/1553374.1553441 · DOI · Paper page
- Lee L., 2019, Efficient exploration via state marginal matching
- 10.1098/rsif.2013.0069 · DOI · Paper page
- Millidge B., 2020, Whence the expected free energy?
- Millidge B., 2020, On the relationship between active inference and control as inference
- Mnih V., 2016, Proceedings of the International Conference on Machine Learning, 1928
- Mnih V., 2013, Playing Atari with deep reinforcement learning
- Mohamed, S. & Rezende, D. J. (2015).Variational information maximisation for intrinsically motivated reinforcement learning. arXiv:1509.08731. · arXiv
- 10.3389/fnhum.2014.00160 · DOI · Paper page
- Ng A. Y., 2003, Shaping and policy search in reinforcement learning
- O'Donoghue B., 2020, Making sense of reinforcement learning and probabilistic inference
- O'Donoghue B., 2018, Proceedings of the International Conference on Machine Learning, 3836
- Osband I., 2016, Advances in neural information processing systems, 29, 4026
- Padakandla S., 2019, Reinforcement learning in non-stationary environments
- 10.1162/neco_a_01102 · DOI · Paper page
- 10.1007/s00213-019-05240-0 · DOI · Paper page
- Pathak D., 2017, Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, 16
- Rakelly K., 2019, Proceedings of the International Conference on Machine Learning, 5331
- Ross S., 2008, Advances in neural information processing systems, 20, 1225
- 10.1080/09540090600768658 · DOI · Paper page
- 10.1016/j.mehy.2014.12.007 · DOI · Paper page
- 10.7554/eLife.41703 · DOI · Paper page
- Sekar R., 2020, Planning to explore via self-supervised world models
- Sorg J., 2012, Variance-based rewards for approximate Bayesian reinforcement learning
- 10.1016/B978-1-55860-141-3.50030-4 · DOI · Paper page
- Sutton R. S., 1998, Introduction to reinforcement learning
- Sutton R. S., 2018, Reinforcement learning: An introduction
- Tijsma A. D., 2016, Proceedings of the IEEE Symposium Series on Computational Intelligence, 1
- Vermorel J., 2005, Proceedings of theEuropean Conference on Machine Learning, 437
- Watkins C. J. C. H., 1989, Learning from delayed rewards
- Wiering M., 1998, Proceedings of the Sixth International Conference on Simulation of Adaptive Behavior: From Animals to Animats, 6, 223
- Zintgraf L., 2019, Varibad: A very good method for Bayes-adaptive deep Rl via meta-learning
Cited by
A step-by-step tutorial on active inference and its application to empirical data
Smith, R., Friston, K. J., & Whyte, C. J. (2022) · Journal of Mathematical Psychology · doi:10.1016/j.jmp.2021.102632 · Cited by 246
Self-citationDeeply Felt Affect: The Emergence of Valence in Deep Active Inference
Hesp, C., Smith, R., Parr, T., Allen, M., Friston, K. J., & Ramstead, M. J. D. (2021) · Neural Computation · doi:10.1162/neco_a_01341 · Cited by 246
“…between reinforcement learning and active inference also tend to provide evidence for the claim that the latter tends to have comparable performance to, or can outperform, the latter, especially in environments with changing contingencies and sparse rewards (Sajid, Ball, & Friston, 2020).”
Self-citationThe free energy principle made simpler but not too simple
Friston, K., Da Costa, L., Sajid, N., Heins, C., Ueltzhöffer, K., Pavliotis, G. A., & Parr, T. (2023) · Physics Reports · doi:10.1016/j.physrep.2023.07.001 · Cited by 163
“It describes active inference based upon autonomous paths or policies and has been used in a variety of applications and simulations; ranging from games in behavioural economics [152] and reinforcement learning [153,154] through to language understanding [155] and scene construction [156].”
Self-citationThe empirical status of predictive coding and active inference
Hodson, R., Mehta, M., & Smith, R. (2024) · Neuroscience & Biobehavioral Reviews · doi:10.1016/j.neubiorev.2023.105473 · Cited by 69
pymdp: A Python library for active inference in discrete state spaces
Heins, C., Millidge, B., Demekas, D., Klein, B., Friston, K., Couzin, I. D., & Tschantz, A. (2022) · Journal of Open Source Software · doi:10.21105/joss.04098 · Cited by 70
Self-citation
Metadata
- Authors
- Sajid, N., Ball, P. J., Parr, T., & Friston, K. J.
- Venue
- Neural Computation
- Publisher
- MIT Press
- Published
- 2021-03
- Volume
- 33
- Issue
- 3
- Pages
- 674-712
- ISSN
- 0899-7667 (ISSN-L)
- 1530-888X
- OpenAlex
- W3003886363
- Cited by
- 145
- References
- 31
- FWCI
- 8.65
- Type
- journal article
- Language
- en
- Open access
- green
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.