AI for Autonomous Ships Challenges in Design and Validation

Transcription

1 VTT TECHNICAL RESEARCH CENTRE OF FINLAND LTD AI for Autonomous Ships Challenges in Design and Validation ISSAV 2018 Eetu Heikkilä

2 Autonomous ships - activities in VTT Autonomous ship systems Unmanned engine room Situation awareness Autonomous autopilot Connectivity Human factors of remote and autonomous systems Safety assessment 22/03/2018 2

3 Contents AI technologies for autonomous shipping Design & validation challenges Methodological Technical Acknowledgements: Olli Saarela, Heli Helaakoski, Heikki Ailisto, Jussi Martio 22/03/2018 3

4 AI technologies

5 Definitions - Autonomy Autonomy is the ability of a system to achieve operational goals in complex domains by making decisions and executing actions on behalf of or in cooperation with humans. (NFA, 2012) Target to increase productivity, cost efficiency, and safety Not only by reducing human work, but also by enabling new business logic Level Name 1 Human operated 2 Human assisted 3 Human delegated 4 Supervised 5 Mixed initiative 6 Fully autonomous 22/03/2018 5

6 Definitions - Artificial Intelligence The Turing test: Can a person tell which of the other parties is a machine AI is a moving target When a computer program is able to perform a task, people start to consider the task merely computational, not requiring actual intelligence after all. AI is all the software we don t yet know how to write. More practically: AI is a collection of technologies facilitating smart operation of machines and systems. Conclusions and actions fitting the prevailing situation. In many cases learning from data or experience. Picture: 22/03/2018 J.A. Sánchez Margallo, Wikipedia 6

7 It s all about better utilization of data Monitor and report in real-time Predict for next best actions Optimize logistics, energy and raw materials Expose new business opportunities Information utilisation Integration Integrate analytics as a part of enterprise IT systems and decision chains Artificial intelligence methods for business purposes Application specific implementation of algorithms and analysis Appropriate tools for domain specific applications Intelligence Visualization Easy-to-understand and descriptive visualization of complex data Interactive methods for focusing relevant parts of the data Ensure reliable and relevant data from multiple sources Data fusion and validation Data acquisition and storage Collect past and real-time data Acquire essential data from different sources Manage high volumes of varying data 22/03/2018 7

8 Stages in AI development 1. Weak AI, Narrow AI Focused on one narrow task, e.g., some game or diagnosis of a particular disease Very limited adaptability, e.g., if the rules of a game are changed even slightly All current AI applications are Weak AI 2. Multi-agent systems Interaction of several weak AI applications The whole is larger than the sum of the parts Being developed, e.g., autonomous vehicles, virtual assistants (Apple's Siri, Amazon Alexa, ) 3. Strong AI, General AI Wide applicability and adaptability Human-like consciousness An evasive long-term research goal 4. Super AI Machine intelligence exceeds human intelligence Singularity: AI develops even more powerful AI Machines might take over. Maybe some day (or some century) 22/03/2018 8

9 Machine learning The bread and butter of the current AI boom, especially Deep learning Reinforcement learning Supervised learning Given x and y data, learn y = f x + e Unsupervised learning Given x data, discover patterns in it Clustering, dimensionality reduction, anomaly detection, Simple f x Model identification Statistical pattern recognition Artificial intelligence Complex f x 22/03/2018 9

10 Deep learning Supervised learning with complex models Especially large Artificial Neural Networks Possibly millions of model parameters identified from data Very good results in complex modelling Nonlinear multivariate models E.g., image classification Downsides Decisions cannot be well explained Complex nonlinear models can behave strangely for some inputs 22/03/2018 Image from Jeff Clune: 10 Deep Learning Overview

11 Reinforcement learning Determine an action based on balancing Exploitation of previous good choices Exploration of possibilities not yet tried Observe the result from the action Action Agent Environment Observation Global optimization that builds a model with possibly a large number of parameters, e.g., a deep neural network. Requires a large number of iterations Games AlphaGo playing against itself Consumer analytics You may also be interested in " Simulation models instead of real processes Random trials on real processes might be dangerous Validity of the simulator? 22/03/

12 Many more techniques are called AI depending on task & model complexity Transfer learning Adapt a large data set from a more-or-less similar task to supplement a small data set available from a new task. Reasoning Rule-based systems, decision trees, case-based reasoning, Evolutionary computation Genetic algorithms, for challenging optimization tasks Translation of natural languages 22/03/

13 Vision and natural language Pictures with 200 categories, e.g. ant Answering natural language questions based on pictures Speech recognition from telephone calls A: they think lunch is too long B: {laugh} A: so they're going to have like %uh thirty minutes for each period and they're going to extend the periods we're going to have more periods B: oh God 22/03/ Y. Shoham et al: AI Index, November 2017

14 Uses for AI in autonomous ships Situational awareness Surroundings Ship systems Decision-making Route planning Navigational decisions 22/03/

15 Design & validation challenges

16 Autonomy and AI vs. safety Introduction of new technologies and ways of working brings along new and modified safety risks Increasing system complexity New interactions between humans and machines Lack of prescriptive standards increases the technology developers responsibility for assuring safety 22/03/

17 Selected challenges in design & validation 22/03/

18 Concept design Focus of development activities shifts towards the early concept design phase Quality of system description Including operating environment, stakeholders, interfaces Concept of operations Requirements management Goals for the system performance 22/03/

19 Architecture & detailed design Reliable handling of large data amounts needs to be ensured Planning of data usage to teach the system 22/03/

20 Data quality issues are often realized only at a very late stage 22/03/2018 Schroeck, M., Shockley, R., Smart, J., Romero-Morales, D. & Tufano, P. (2012). Analytics: 20 The real-world use of big data. IBM Global Services.

21 Implementation & integration How to ensure the system learns the right things? W-model Increasing need for simulator testing Transparency of machine learning Training 22/03/

22 Models can behave strangely for some inputs Panda <1% distortion Gibbon (99.3% confidence) + = Distortions can be crafted to produce the desired erroneous outcome Example from 22/03/

23 Verification & Validation Lack of prescriptive standards Technology developer increasingly responsible for demonstrating the safety Goal-based approach used to link safety evidence & system requirements 22/03/

24 V&V methodology: Goal-based approach Problem: How to create a comprehensible link between the safety goals and evidence? System modeled as a structure of safety goals GSN argumentation G0 modeling Top Goal language GOAL Is solved by Is solved by G1 Goal G2 Goal GOAL Is solved by Is solved by GOAL G1.1 Sub-Goal G1.2 Sub-Goal GOAL GOAL 22/03/

25 Operation Change management New operational logic, increased human-machine interaction 22/03/

26 Conclusions

27 Conclusions AI technologies bring both opportunities and risks in the maritime sector Robust process for V&V of AI systems is needed Domain understanding needs to be incorporated in all stages of development 22/03/

28 TECHNOLOGY FOR BUSINESS Contact information: