Jan Gruteser (talk | contribs) m (→SimB) |
|||
Line 1: | Line 1: | ||
== SimB == | == SimB == | ||
Additional documentation is available here: [https://github.com/hhu-stups/prob2_ui/blob/develop/src/main/helpsources/help_en/Main%20Menu/Advanced/SimB.md SimB Documentation] | |||
SimB is a simulator built on top of ProB. It is available in the latest SNAPSHOT version in the new JavaFX based user interface [[ProB2-UI]] (https://github.com/hhu-stups/prob2_ui), The modeler can write SimB annotations for a formal model to simulate it. Examples are available at https://github.com/favu100/SimB-examples. | SimB is a simulator built on top of ProB. It is available in the latest SNAPSHOT version in the new JavaFX based user interface [[ProB2-UI]] (https://github.com/hhu-stups/prob2_ui), The modeler can write SimB annotations for a formal model to simulate it. Examples are available at https://github.com/favu100/SimB-examples. |
Additional documentation is available here: SimB Documentation
SimB is a simulator built on top of ProB. It is available in the latest SNAPSHOT version in the new JavaFX based user interface ProB2-UI (https://github.com/hhu-stups/prob2_ui), The modeler can write SimB annotations for a formal model to simulate it. Examples are available at https://github.com/favu100/SimB-examples. Furthermore, it is then possible to validate probabilistic and timing properties with statistical validation techniques such as hypothesis testing and estimation.
SimB also contains a feature called interactive simulation. This feature allows user interaction to trigger a simulation. For interactive simulation, a modeler has to encode SimB listeners on events, triggering a SimB simulation. Interactive Simulation examples are available at https://github.com/favu100/SimB-examples/tree/main/Interactive_Examples.
More recently, SimB is extended by a new feature which makes it possible to load an Reinforcement learning Agent.
Technically, each step of the RL agent is converted into a SimB activation.
In order to simulate a RL agent in SimB, one must (1) create a formal B model for the RL agent, and (2) create a mapping between the state in the RL agent and the formal model, and (3) provide information to the formal B model again.
Reinforcement Learning examples are available at: https://github.com/hhu-stups/reinforcement-learning-b-models
To cite SimB as a tool, its timing or probabilistic simulation features, or SimBs statistical validation techniques, please use:
@InProceedings{simb, Author = {Vu, Fabian and Leuschel, Michael and Mashkoor, Atif}, Title = {{Validation of Formal Models by Timed Probabilistic Simulation}}, Booktitle = {Proceedings ABZ}, Year = 2021, Series = {LNCS}, Volume = {12709}, Pages = {81--96} }
To cite SimBs interactive simulation, please use:
@InProceedings{simb, Author = {Vu, Fabian and Leuschel, Michael}, Title = {{Validation of Formal Models by Interactive Simulation}}, Booktitle = {Proceedings ABZ}, Year = 2023 Series = {LNCS}, Volume = {14010}, Pages = {59--69} }
Start SimB via "SimB" in the "Advanced" Menu after opening a machine.
Now, you can open a SimB file (JSON format) controlling the underlying formal model.
A SimB file consists of SimB activations and SimB listeners to simulate the model. SimB activations encode an activation diagram with probabilistic and timing elements for automatic simulation. To enable interactive simulation, it is also necessary to encode interactive elements aka. SimB listeners which trigger other SimB activations. Within these elements, the modeler can user B expressions which are evaluated on the current state.
The general structure of a SimB simulation is as follows:
{ "activations": [...] "listeners": [...] }
activations stores SimB activations, while listeners stores SimB listeners. activations must be encoded, listeners is optional and defaults to the empty list.
The SimB file always contains an activations field storing a list of probabilistic and timing elements to control the simulation. Probabilistic values are always evaluated to ensure that the sum of all possibilities is always 1. Otherwise an error will be thrown at runtime. There are two types of activations: direct activation and probabilistic choice. All activations are identified by their id.
A direct activation activates an event to be executed in the future. It requires the fields id, and execute to be defined. All other fields can be defined optionally.
{ "id": ... "execute": ... "after": ... "activating": ... "activationKind": ... "additionalGuards": ... "fixedVariables": .... "probabilisticVariables": .... "priority": ... }
A probabilistic choice selects an event to be executed in the future. It requires the two fields id, and chooseActivation. chooseActivation is a Map storing Key-Value pairs where activations (identified by their id) are mapped to a probability. It is possible to chain multiple probabilistic choices together, but eventually, a direct activation must be reached.
{ "id": ... "chooseActivation": ... }
In the following, an example for a SimB file controlling a Traffic Lights for cars and pedestrians (with timing and probabilistic behavior) is shown:
{ "activations": [ {"id":"$initialise_machine", "execute":"$initialise_machine", "activating":"choose"}, {"id":"choose", "chooseActivation":{"cars_ry": "0.8", "peds_g": "0.2"}}, {"id":"cars_ry", "execute":"cars_ry", "after":5000, "activating":"cars_g"}, {"id":"cars_g", "execute":"cars_g", "after":500, "activating":"cars_y"}, {"id":"cars_y", "execute":"cars_y", "after":5000, "activating":"cars_r"}, {"id":"cars_r", "execute":"cars_r", "after":500, "activating":"choose"}, {"id":"peds_g", "execute":"peds_g", "after":5000, "activating":"peds_r"}, {"id":"peds_r", "execute":"peds_r", "after":5000, "activating":"choose"} ] }
Interactive elements in SimB are so called SimB listeners. A SimB listener consists of four fields id, event, predicate, and activating. This means that the SimB listener associated with id listens on a manual/user interaction on event with predicate after which activations in activating are triggered. predicate is optional and defaults to 1=1. Manual/User interaction is recognized via VisB and ProB's Operations View.
{ "id": ... "event": ... "predicate": ... "activating": [...] }
In the following, we parts of a SimB simulation from an automotive case study. The SimB simulation contains a SimB listener which is linked to two activations. The case study models the car's lighting system controlled by pitman controller, the key ignition, and the warning lights button.
The SimB listeners states that SimB listens on user interaction on the event ENV_Pitman_DirectionBlinking, to trigger two SimB activations blinking_on and blinking_off afterward. Practically, this means that a driver's input on the pitman for direction blinking activates the blinking cycle for the corresponding direction indicators.
{ "activations": [ { "id": "blinking_on", "execute": "RTIME_BlinkerOn", "after": "curDeadlines(blink_deadline)", "activating" : "blinking_off", ... }, { "id": "blinking_off", "execute": "RTIME_BlinkerOff", "after": "curDeadlines(blink_deadline)", "activating" : "blinking_on", ... }, ... ] "listeners": [ { "id": "start_blinking", "event": "ENV_Pitman_DirectionBlinking", "activating" : ["blinking_on", "blinking_off"] } ] }
Instead of loading a SimB simulation as a JSON file, one can also load a Reinforcement Learning agent implemented in Python (.py). For the simulation to work, one has to apply the following steps:
1. Train a model of a Reinforcement Learning agent.
2. Create a formal B model (including safety shield) for the RL agent.
The operations represent the actions the RL agent can choose from. The formal model's state mainly represents the state of the environment. Safety shields are encoded by the operations' guards which are provided to the RL agent. Enabled operations are considered to be safe. Thus, the RL agent chooses the enabled operation/action with the highest predicted reward. The operations' substitutions model the desired behavior of the respective actions.
An example for the FASTER of a HighwayEnvironment is as follows:
FASTER = PRE ¬(∃v. (v ∈ PresentVehicles \ {EgoVehicle} ∧ VehiclesX(v) > 0.0 ∧ VehiclesX(v) < 45.0 ∧ VehiclesY(v) < 3.5 ∧ VehiclesY(v) > -3.5)) THEN Crash :∈ BOOL || PresentVehicles :∈ P(Vehicles) ; VehiclesX :∈ Vehicles → R || VehiclesY :∈ Vehicles → R || VehiclesVx :| (VehiclesVx ∈ PresentVehicles → R ∧ VehiclesVx(EgoVehicle) ≥ VehiclesVx’(EgoVehicle) - 0.05) || VehiclesVy :∈ Vehicles → R || VehiclesAx :| (VehiclesAx ∈ PresentVehicles → R ∧ VehiclesAx(EgoVehicle) ≥ -0.05) || VehiclesAy :∈ Vehicles → R || Reward :∈ R END
Lines 2-4 shows the operation's guard which is used as safety shield.
Lines 6-15 shows the operation's substitution describing the desired behavior after executing FASTER.
3. Implement the mapping between the RL agent in Python and the formal B model. This includes the mapping of actions to operations, and the mapping of information from the RL agent, particularly the environment and observation, to the variables.
An example for the mapping of actions to operations is as follows:
action_names = { 0: "LANE_LEFT", 1: "IDLE", 2: "LANE_RIGHT", 3: "FASTER", 4: "SLOWER" }
Here, one can see that the numbers representing the actions in the RL agents are mapped to the corresponding operation names in the formal B model.
An example for a mapping to a variable in the formal B model is as follows:
def get_VehiclesX(obs): return "{{EgoVehicle |-> {0}, Vehicles2 |-> {1}, Vehicles3 |-> {2}, Vehicles4 |-> {3}, Vehicles5 |-> {4}}}" .format(obs[0][1]*200, obs[1][1]*200, obs[2][1]*200, obs[3][1]*200, obs[4][1]*200) # Implemented manually
Remark: While the getter for the variable is generated by B2Program, the function for the mapping is implemented manually.
4. Implement necessary messages sent between the ProB animator and the RL agent. Simulation should be a while-loop which runs while the simulation has not finished.
Example code (line 70 - 113; particularly 86 - 113): https://github.com/hhu-stups/reinforcement-learning-b-models/blob/main/HighwayEnvironment/HighwayEnvironment.py
To generate a RL agent for SimB, one can use the high-level code generator B2Program: https://github.com/favu100/b2program
Given a formal B model, B2Program generates an RL agent which loads a given trained model and execute the necessary steps. This includes the fourth step described before. The third step still has to be implemented; in this step, B2Program only generates the templates for the mappings which are then completed manually.
Using a SimB file, the modeler can play a single simulation on the underlying model in real-time. The modeler can then manually check whether the model behaves as desired. Combining VisB and SimB, a simulation can be seen as an animated picture similar to a GIF picture. This gives the domain expert even a better understanding of the model. With SimB listeners, it is also possible to encode simulations that are triggered by manual/user actions.
A Traffic Light example (based on the SimB file shown in Using_SimB ) simulating the first 21 seconds is shown below.
Based on a single simulation, the modeler can generate a timed trace which is also stored in the SimB format. Afterwards, it can be used to replay a scenario. similar to real-time simulation.
Below, the resulting timed trace for the scenario in Real-Time Simulation is shown
{ "activations": [ { "execute": "$initialise_machine", "after": "0", "priority": 0, "additionalGuards": null, "activationKind": null, "fixedVariables": { "tl_cars": "red", "tl_peds": "red" }, "probabilisticVariables": null, "activating": [ "cars_ry_1" ], "id": "$initialise_machine" }, { "execute": "cars_ry", "after": "5000", "priority": 0, "additionalGuards": null, "activationKind": null, "fixedVariables": null, "probabilisticVariables": null, "activating": [ "cars_g_2" ], "id": "cars_ry_1" }, { "execute": "cars_g", "after": "500", "priority": 0, "additionalGuards": null, "activationKind": null, "fixedVariables": null, "probabilisticVariables": null, "activating": [ "cars_y_3" ], "id": "cars_g_2" }, { "execute": "cars_y", "after": "5000", "priority": 0, "additionalGuards": null, "activationKind": null, "fixedVariables": null, "probabilisticVariables": null, "activating": [ "cars_r_4" ], "id": "cars_y_3" }, { "execute": "cars_r", "after": "500", "priority": 0, "additionalGuards": null, "activationKind": null, "fixedVariables": null, "probabilisticVariables": null, "activating": [ "peds_g_5" ], "id": "cars_r_4" }, { "execute": "peds_g", "after": "5000", "priority": 0, "additionalGuards": null, "activationKind": null, "fixedVariables": null, "probabilisticVariables": null, "activating": [ "peds_r_6" ], "id": "peds_g_5" }, { "execute": "peds_r", "after": "5000", "priority": 0, "additionalGuards": null, "activationKind": null, "fixedVariables": null, "probabilisticVariables": null, "activating": null, "id": "peds_r_6" } ], "metadata": { "fileType": "Timed_Trace", "formatVersion": 1, "savedAt": "2021-03-03T11:04:08.460477Z", "creator": "User", "proB2KernelVersion": "4.0.0-SNAPSHOT", "proBCliVersion": null, "modelName": null } }
It is also possible to apply Monte Carlo simulation to generate a certain number of simulations. Here, all simulations are played without real time. However, it is possible for the user, to replay the generated scenarios with real-time afterwards.
The input parameters are:
Furthermore, there are two statistical validation techniques that can be applied based on Monte Carlo simulations: Hypothesis Testing and Estimation.
Hypothesis Testing expects the same parameters as Monte Carlo Simulation: Max Steps before Simulation, Number of Simulations, Starting Condition and Ending Condition.
The additional input parameters for Hypothesis Testing are:
Estimation expects the same parameters as Monte Carlo Simulation: Number of Simulations, Starting Condition and Ending Condition.
The additional input parameters for Estimation are:
For an estimated value e, a desired value d, and an epsilon eps, it checks for each simulation whether e is within [d - eps, d + eps]