Abstract
Tangled Program Graphs (TPG) represents a genetic programming framework in which emergent modularity incrementally composes programs into teams of programs into graphs of teams of programs. To date, the framework has been demonstrated on reinforcement learning tasks with stochastic partially observable state spaces or time series prediction. However, evolving solutions to reinforcement tasks often requires agents to demonstrate/ juggle multiple properties simultaneously. Hence, we are interesting in maintaining a population of diverse agents. Specifically, agent performance on a reinforcement learning task controls how much of the task they are exposed to. Premature convergence might therefore preclude solving aspects of a task that the agent only later encounters. Moreover, ‘pointless complexity’ may also result in which graphs largely consist of hitchhikers. In this research we benchmark the utilization of rampant mutation (multiple mutations applied simultaneously for offspring creation) and action programs (multiple actions per state). Several parameterizations are also introduced that potentially penalize the introduction of hitchhikers. Benchmarking over five VizDoom tasks demonstrates that rampant mutation reduces the likelihood of encountering pathologically bad offspring while action programs appears to improve performance in four out of five tasks. Finally, use of TPG parameterizations that actively limit the complexity of solutions appears to result in very efficient low dimensional solutions that generalize best across all combinations of 3, 4 and 5 VizDoom tasks.
Access this chapter
Tax calculation will be finalised at checkout
Purchases are for personal use only
Notes
- 1.
Implies that the interaction represents the special case of an episodic task [24].
- 2.
Although a minimum of two learners (with different actions) is necessary to avoid defining a degenerate team Sect. 1.2.2.
- 3.
An arc marking scheme has since been proposed [9], however, for the purpose of this work the original team formulation was assumed.
- 4.
Stochastic nature of each subtask requires that agents are evaluated over multiple initializations.
- 5.
- 6.
Reflected in the parameterization of the ‘Rampant Magnitude’ row in Table 1.1.
- 7.
Includes introns and hitchhikers.
References
Bjedov, I., Tenaillon, O., Gerard, B., Souza, V., Denamur, E., Radman, M., Taddei, F., Matic, I.: Stress-induced mutagenesis in bacteria. Science 300, 1404–1409 (2003)
Brameier, M., Banzhaf, W.: Linear Genetic Programming. Springer (2007)
Branke, J.: Evolutionary approaches to dynamic environments—a survey. In: GECCO Workshop on Dynamic Optimization Problems, pp. 134–137 (1999)
Cobb, H.G.: An investigation into the use of hypermutation as an adaptive operating in genetic algorithms having continuous, time-dependent non-stationary environments. Technical Report TR AIC-90-001, Naval research Laboratory (1990)
Demsar, J.: Statistical comparisons of classifiers over multiple data sets. J. Mach. Learn. Res. 7, 1–30 (2006)
Ghosh, A., Tstutsui, S., Tanaka, H.: Function optimization in non-stationary environment using steady state genetic algorithms with aging of individuals. In: IEEE Congress on Evolutionary Computation, pp. 666–671 (1998)
Grefenstette, J.J.: Genetic algorithms for changing environments. In: PPSN, pp. 137–144 (1992)
Hwangbo, J., Lee, J., Dosovitskiy, A., Bellicoso, D., Tsounis, V., Koltun, V., Hutter, M.: Learning agile and dynamic motor skills for legged robots. CoRR (2019). arXiv:abs/1901.08652
Ianta, A., Amaral, R., Bayer, C., Smith, R.J., Heywood, M.I.: On the impact of tangled program graph marking schemes under the atari reinforcement learning benchmark. In: Proceedings of the ACM Genetic and Evolutionary Computation Conference, p. to appear (2021)
Jaderberg, M., Czarnecki, W.M., Dunning, I., Marris, L., Lever, G., Castañeda, A.G., Beattie, C., Rabinowitz, N.C., Morcos, A.S., Ruderman, A., Sonnerat, N., Green, T., Deason, L., Leibo, J.Z., Silver, D., Hassabis, D., Kavukcuoglu, K., Graepel, T.: Human-level performance in 3D multiplayer games with population-based reinforcement learning. Science 364, 859–865 (2019)
Kelly, S., Heywood, M.I.: Emergent tangled graph representations for atari game playing agents. In: European Conference on Genetic Programming, LNCS, vol. 10196, pp. 64–79 (2017)
Kelly, S., Heywood, M.I.: Emergent solutions to high-dimensional multitask reinforcement learning. Evol. Comput. 26(3), 347–380 (2018)
Kelly, S., Newsted, J., Banzhaf, W., Gondro, C.: A modular memory framework for time series prediction. In: Proceedings of the ACM Genetic and Evolutionary Computation Conference, pp. 949–957 (2020)
Kelly, S., Smith, R.J., Heywood, M.I.: Emergent policy discovery for visual reinforcement learning through tangled program graphs: a tutorial. In: Banzhaf, W., Spector, L., Sheneman L (eds.) Genetic Programming Theory and Practice XVI, Genetic and Evolutionary Computation, pp. 37–57 (2018)
Kelly, S., Smith, R.J., Heywood, M.I., Banzhaf, W.: Emergent tangled program graphs in partially observable recursive forecasting and ViZDoom navigation tasks. ACM Trans. Evol. Learn. Optim. 1 (2021)
Kempka, M., Wydmuch, M., Runc, G., Toczek, J., Jaskowski, W.: ViZDoom: A Doom-based AI research platform for visual reinforcement learning. In: IEEE Conference on Computational Intelligence and Games, pp. 1–8 (2016)
Koza, J.R.: Genetic Programming—On the Programming of Computers by Means of Natural Selection. MIT Press, Complex Adaptive Systems (1993)
Moriarty, D.E., Schultz, A.C., Grefenstette, J.J.: Evolutionary algorithms for reinforcement learning. J. Artif. Intell. Res. 11, 199–229 (1999)
Parter, M., Kashtan, N., Alon, U.: Facilitated variation: how evolution learns from past environments to generalize to new environments. PLOS Comput. Biol. 4(11), 1–15 (2008)
Smith, R.J., Heywood, M.I.: Scaling tangled program graphs to visual reinforcement learning in ViZDoom. In: European Conference on Genetic Programming, Lecture LNCS, vol. 10781, pp. 135–150 (2018)
Smith, R.J., Heywood, M.I.: Evolving Dota 2 shadow fiend bots using genetic programming with external memory. In: Proceedings of the ACM Genetic and Evolutionary Computation Conference, pp. 179–187 (2019)
Smith, R.J., Heywood, M.I.: A model of external memory for navigation in partially observable visual reinforcement learning tasks. In: European Conference on Genetic Programming, LNCS, vol. 11451, pp. 162–177 (2019)
Sünderhauf, N., Brock, O., Scheirer, W.J., Hadsell, R., Fox, D., Leitner, J., Upcroft, B., Abbeel, P., Burgard, W., Milford, M., Corke, P.: The limits and potentials of deep learning for robotics. Int. J. Robot. Res. 37(4–5), 405–420 (2018)
Sutton, R.S., Barto, A.G.: Reinforcement Learning: An Introduction. MIT (2018)
Teng, G., Popavasiliou, F.N.: Immunoglobulin somatic hypermutation. Annu. Rev. Genet. 41, 107–120 (2007)
Acknowledgements
We gratefully acknowledge support from the NSERC CRD and Discovery programs (Canada).
Author information
Authors and Affiliations
Corresponding author
Editor information
Editors and Affiliations
Rights and permissions
Copyright information
© 2022 The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd.
About this chapter
Cite this chapter
Bayer, C., Amaral, R., Smith, R.J., Ianta, A., Heywood, M.I. (2022). Finding Simple Solutions to Multi-Task Visual Reinforcement Learning Problems with Tangled Program Graphs. In: Banzhaf, W., Trujillo, L., Winkler, S., Worzel, B. (eds) Genetic Programming Theory and Practice XVIII. Genetic and Evolutionary Computation. Springer, Singapore. https://doi.org/10.1007/978-981-16-8113-4_1
Download citation
DOI: https://doi.org/10.1007/978-981-16-8113-4_1
Published:
Publisher Name: Springer, Singapore
Print ISBN: 978-981-16-8112-7
Online ISBN: 978-981-16-8113-4
eBook Packages: Computer ScienceComputer Science (R0)