Skip to content

Commit

Permalink
Fixed merge conflict
Browse files Browse the repository at this point in the history
  • Loading branch information
bkushigian committed Aug 20, 2024
2 parents 1e0067b + aa10c4c commit 18f6d60
Show file tree
Hide file tree
Showing 10 changed files with 588 additions and 82 deletions.
200 changes: 200 additions & 0 deletions DESIGN.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,200 @@
# Design

This document is a description, as far as I understand it, of the inner design
of the solver and PostFlopGame. This is a working document for me to get my
bearings.

## PostFlopGame

### Build/Allocate/Initialize

To set up a `PostFlopGame` we need to **create a `PostFlopGame` instance**,
**allocate global storage and `PostFlopNode`s**, and **initialize the
`PostFlopNode` child/parent relationship**. This is done in several steps.

We begin by creating a `PostFlopGame` instance.

```rust
let mut game = PostFlopGame::with_config(card_config, action_tree).unwrap();
```

A `PostFlopGame` requires an
`ActionTree` which describes all possible actions and lines (no runout
information), and a `CardConfig`, which describes player ranges and
flop/turn/river data.

Once we have created a `PostFlopGame` instance we need to allocate the following
memory and initialize its values:

+ `game.node_arena`
+ `game.storage1`
+ `game.storage2`
+ `game.storage_ip`
+ `game.storage_chance`

These fields are not allocated/initialized at the same time:

+ `game.node_arena` is allocated and initialized via `with_config()` (i.e., when
we created our `PostFlopGame`),
+ other storage is allocated via `game.allocate_memory()`.

#### Allocating and Initializing `node_arena`

We constructed a `PostFlopGame` by calling
`PostFlopGame::with_config(card_config, action_tree)`, which under the hood
actually calls:

```rust
let mut game = Self::new();
game.update_config(card_config, action_tree)?;
```

`PostFlopGame::update_config` sets up configuration data, sanity checks things
are correct, and then calls `self.init_root()`.

`init_root` is responsible for:

1. Counting number of `PostFlopNode`s to be allocated (`self.nodes_per_street`),
broken up by flop, turn, and river
2. Allocating `PostFlopNode`s in the `node_arena` field
3. Clearing storage: `self.clear_storage()` sets each storage item to a new
`Vec`
4. Invoking `build_tree_recursive` which initializes each node's child/parent
relationship via `child_offset` (through calls to `push_actions` and
`push_chances`).

Each `PostFlopNode` points to node-specific data (e.g., strategies and
cfregrets) that is located inside of `PostFlopGame.storage*` fields (which is
currently unallocated) via similarly named fields `PostFlopNode.storage*`.

Additionally, each node points to the children offset with `children_offset`,
which records where in `node_arena` relative to the current node that node's
children begin. We allocate this memory via:

```rust
game.allocate_memory(false); // pass `true` to use compressed memory
```

This allocates the following memory:

+ `self.storage1`
+ `self.storage2`
+ `self.storage3`
+ `self.storage_chance`

Next, `allocate_memory()` calls `allocate_memory_nodes(&mut self)`, which
iterates through each node in `node_arena` and sets storage pointers.

After `allocate_memory` returns we still need to set `child_offset`s.

### Storage

There are several fields marked as `// global storage` in `game::mod::PostFlopGame`:

```rust
// global storage
// `storage*` are used as a global storage and are referenced by `PostFlopNode::storage*`.
// Methods like `PostFlopNode::strategy` define how the storage is used.
node_arena: Vec<MutexLike<PostFlopNode>>,
storage1: Vec<u8>,
storage2: Vec<u8>,
storage_ip: Vec<u8>,
storage_chance: Vec<u8>,
locking_strategy: BTreeMap<usize, Vec<f32>>,
```

These are referenced from `PostFlopNode`:

```rust
storage1: *mut u8, // strategy
storage2: *mut u8, // regrets or cfvalues
storage3: *mut u8, // IP cfvalues
```

+ `storage1` seems to store the strategy
+ `storage2` seems to store regrets/cfvalues, and
+ `storage3` stores IP's cf values (does that make `storage2` store OOP's cfvalues?)

Storage is a byte vector `Vec<u8>`, and these store floating point values.

> [!IMPORTANT]
> Why are these stored as `Vec<u8>`s? Is this for swapping between
> `f16` and `f32`s?
Some storage is allocated in `game::base::allocate_memory`:

```rust
let storage_bytes = (num_bytes * self.num_storage) as usize;
let storage_ip_bytes = (num_bytes * self.num_storage_ip) as usize;
let storage_chance_bytes = (num_bytes * self.num_storage_chance) as usize;

self.storage1 = vec![0; storage_bytes];
self.storage2 = vec![0; storage_bytes];
self.storage_ip = vec![0; storage_ip_bytes];
self.storage_chance = vec![0; storage_chance_bytes];
```

`node_arena` is allocated in `game::base::init_root()`:

```rust
let num_nodes = self.count_nodes_per_street();
let total_num_nodes = num_nodes[0] + num_nodes[1] + num_nodes[2];

if total_num_nodes > u32::MAX as u64
|| mem::size_of::<PostFlopNode>() as u64 * total_num_nodes > isize::MAX as u64
{
return Err("Too many nodes".to_string());
}

self.num_nodes = num_nodes;
self.node_arena = (0..total_num_nodes)
.map(|_| MutexLike::new(PostFlopNode::default()))
.collect::<Vec<_>>();
self.clear_storage();
```

`locking_strategy` maps node indexes (`PostFlopGame::node_index`) to a locked
strategy. `locking_strategy` is initialized to an empty `BTreeMap<usize,
Vec<f32>>` by deriving Default. It is inserted into via
`PostFlopGame::lock_current_strategy`

### Serialization/Deserialization

Serialization relies on the `bincode` library's `Encode` and `Decode`. We can set
the `target_storage_mode` to allow for a non-full save. For instance,

```rust
game.set_target_storage_mode(BoardState::Turn);
```

will ensure that when `game` is encoded, it will only save Flop and Turn data.
When a serialized tree is deserialized, if it is a partial save (e.g., a Turn
save) you will not be able to navigate to unsaved streets.

Several things break when we deserialize a partial save:

+ `node_arena` is only partially populated
+ `node.children()` points to raw data when `node` points to an street that is
not serialized (e.g., a chance node before the river for a Turn save).

### Allocating `node_arena`

We want to first allocate nodes for `node_arena`, and then run some form of
`build_tree_recursive`. This assumes that `node_arena` is already allocated, and
recursively visits children of nodes and modifies them to

### Data Coupling/Relations/Invariants

+ A node is locked IFF it is contained in the game's locking_strategy
+ `PostFlopGame.node_arena` is pointed to by `PostFlopNode.children_offset`. For
instance, this is the basic definition of the `PostFlopNode.children()`
function:

```rust
slice::from_raw_parts(
self_ptr.add(self.children_offset as usize),
self.num_children as usize,
)
```

We get a pointer to `self` and add children offset.
53 changes: 53 additions & 0 deletions examples/simple.rs
Original file line number Diff line number Diff line change
@@ -0,0 +1,53 @@
use postflop_solver::*;

fn main() {
// ranges of OOP and IP in string format
// see the documentation of `Range` for more details about the format
let oop_range = "66+";
let ip_range = "66+";

let card_config = CardConfig {
range: [oop_range.parse().unwrap(), ip_range.parse().unwrap()],
flop: flop_from_str("Td9d6h").unwrap(),
turn: NOT_DEALT,
river: NOT_DEALT,
};

// bet sizes -> 60% of the pot, geometric size, and all-in
// raise sizes -> 2.5x of the previous bet
// see the documentation of `BetSizeOptions` for more details
let bet_sizes = BetSizeOptions::try_from(("100%", "100%")).unwrap();

let tree_config = TreeConfig {
initial_state: BoardState::Flop, // must match `card_config`
starting_pot: 200,
effective_stack: 200,
rake_rate: 0.0,
rake_cap: 0.0,
flop_bet_sizes: [bet_sizes.clone(), bet_sizes.clone()], // [OOP, IP]
turn_bet_sizes: [bet_sizes.clone(), bet_sizes.clone()],
river_bet_sizes: [bet_sizes.clone(), bet_sizes],
turn_donk_sizes: None, // use default bet sizes
river_donk_sizes: Some(DonkSizeOptions::try_from("100%").unwrap()),
add_allin_threshold: 1.5, // add all-in if (maximum bet size) <= 1.5x pot
force_allin_threshold: 0.15, // force all-in if (SPR after the opponent's call) <= 0.15
merging_threshold: 0.1,
};

// build the game tree
// `ActionTree` can be edited manually after construction
let action_tree = ActionTree::new(tree_config).unwrap();
let mut game = PostFlopGame::with_config(card_config, action_tree).unwrap();

// allocate memory without compression (use 32-bit float)
game.allocate_memory(false);

// solve the game
let max_num_iterations = 20;
let target_exploitability = game.tree_config().starting_pot as f32 * 0.100; // 10.0% of the pot
let exploitability = solve(&mut game, max_num_iterations, target_exploitability, true);
println!("Exploitability: {:.2}", exploitability);

// get equity and EV of a specific hand
game.cache_normalized_weights();
}
2 changes: 2 additions & 0 deletions src/card.rs
Original file line number Diff line number Diff line change
Expand Up @@ -11,6 +11,8 @@ use bincode::{Decode, Encode};
/// - `card_id = 4 * rank + suit` (where `0 <= card_id < 52`)
/// - `rank`: 2 => `0`, 3 => `1`, 4 => `2`, ..., A => `12`
/// - `suit`: club => `0`, diamond => `1`, heart => `2`, spade => `3`
///
/// An undealt card is represented by Card::MAX (see `NOT_DEALT`).
pub type Card = u8;

/// Constant representing that the card is not yet dealt.
Expand Down
Loading

0 comments on commit 18f6d60

Please sign in to comment.